Disable zexy/lpt on Windows for now, it doesn't work there at present and causes a segfault in the regression tests.

This adds support for the SMMF MIDI message format,
cf. https://bitbucket.org/agraef/pd-smmf.

It also adds canvas/path search for soundfont files, so that these can
be found more easily without having to use absolute pathnames, as well
as a few cosmetic changes to help and error messages.

In default "legacy" mode, the object is 100% backwards-compatible
(except for the soundfont file search), so that existing patches will
continue to work without any changes.

In SMMF mode, which is invoked with the new -smmf option (used either as
a creation argument, or with the "init" message), the SMMF message
format can be used to denote MIDI messages. Legacy messages still
continue to work in SMMF mode as well, except the "note" and "bend"
messages which are used in both formats with the same argument count but
different argument order. However, the corresponding shortcuts "n" and
"b" of the legacy format still work even in SMMF mode.

Rationale: SMMF offers the following advantages over the legacy message
interface that fluid~ currently uses:

- It covers all MIDI voice messages, as well as system exclusive
  messages. (Note that sysex support is particularly interesting in the
  context of fluidsynth because it allows to pass tuning data using
  sysex messages following the MIDI Tuning Standard (MTS), which
  fluidsynth readily supports.)

- It enforces a close 1-1 correspondence between the Pd MIDI objects and
  the message format (it uses the same basenames as message selectors,
  and has the arguments in the right order to easily interface with the
  Pd MIDI objects).

- It is is readily supported by some helper abstractions (midi-input.pd
  and midi-output.pd, available at https://bitbucket.org/agraef/pd-smmf).

- Last but not least, it is compatible with pd-faust and pd-faustgen2
  which makes it very easy to integrate Faust- and soundfont-based
  synthesis in Pd.

* Used to have no margin on the right. Now it does and it matches the left side.

This reverts commit 5eb67db0, reversing
changes made to b363df22.

* Addresses conflict that has emerged due to prior gatom commit.

# Conflicts:
#   pd/src/g_text.c

* Now the shift press is dynamic during the dragging and the grab is properly disengaged after letting go of the mouse, regardless if this happens over the object or outside it.

* gatom due to its reliance on the "click" legacy message lacks adequate information to allow for some of the advanced functionality of the iemgui numbox. This, coupled with other legacy behavior (e.g. having a min and max that does not apply to the values coming into its inlet) in my view make vanilla gatom a legacy object. This also makes a distinguishing factor between it and the iemgui numbox.

* Nonetheless, new features persist, such as drag and shift+drag that offers +-1 and +-0.01 increments respectively inside the number gatom, the ability to append to the symbol gatom text by shift+clicking onto it, and shift+delete deletes all the text at once inside the gatom symbol.

* Reenabled the mouseup event that was mysteriously disabled for an unknown amount of time. This made it difficult to report to the object mouse release event after it has been dragged and the mouse ended-up being positioned outside object's bounds.

* This now allows for very nice control over objects that are being grabbed.

* Mouse release sets doubleclick flag to -1 to differentiate itself from the regular mouse movement which also reports mouse press as 0.

* Allowed for glist_grab to be exclusive in respect to key presses and settable as such via the glist_grab call. See iemgui numbox for benefits of this implementation, described below.

* Completely reworked numbox logic with additional variables designed to improve code readability and disambiguate the x->x_gui.x_changed and x->x_gui.x_change.

* Rotated the order of event propagation to have glist_grab 1st and event bound objects second (e.g. key/keyname etc.). THIS IMPLEMENTATION SHOULD BE HEAVILY SCRUTINIZED TO MAKE SURE THERE ARE NO ADVERSE EFFECTS. Given the rapid release cycle I suggest adding it to the next release and then observing user feedback.

* The numbox now offers the following behavior:

1) click and drag and release (with or without shift which can be pressed at any time to toggle between fine +-0.01 and +-1 increments) which immediately changes the number value and outputs it. This kind of grab does not do exclusive keyboard grab and therefore allows keyboard to still propagate to other bound events.

2) click and immediately release enters the exclusive keyboard mode and makes the '>' appear to make the user aware of this. Here you can use arrows up/down with and without shift, press allowed keyboard keys (numbers, enter, delete, etc.) and edit the value accordingly. Pressing arrows always changes the last digit of the currently truncated number. Pressing enter commits the value and retains the focus for another 3 seconds unless user presses Esc or clicks outside the number box. If all the digits are erased the number implicity assumes 0.

3) clicking and immediately releasing without shift pressed to activate exclusive keyboard mode activates legacy/default behavior where entered text overwrites the existing (the cursor starts at the beginning and initially only has '>'). Doing the same with shift key puts the numbox in the append mode where one can continue adding to the previously committed number. This is even true if the displayed number is simply '+' or '-' due to it exceeding the number box width (not including decimals).

4) Values whenever committed are clipped to the object's min and max and reflected as such in the number object. If the object remains activated, the number is not clipped if it exceeds the object's length, but shows only last n visible digits. Upon removing focus (activation), it redraws from the beginning (if number of digits exceeds the width but does not exceed the min/max thresholds. Re-activating the object redraws it with the last digits visible.

* All glist_grab instances in externals should be taken care of and are, by default, not exclusive.

* Still TODO: checking if vanilla numbox is working ok. The rest of the objects shoould be unaffected.