Only use IUP representation when more compact

[cmyr]

This fixes #568; when presented with deltas that include interpolatable values, we will generate both representations and then choose the more compact.

This is mostly straightforward, but there is some API to bikeshed, particularly around the iup_delta_optimize API.

Previously, this returned the deltas as Vec<Option<Vec2>>, with None values representing deltas that could be interpolated. This meant that we did not have access to these deltas, which are necessary in order for us to generate the non-IUP representation.

We have some options here. The simplest one would be to return the GlyphDelta type that we use when building gvar; the only concern here is that this type represents a point as(i16, i16), where we are currently using Vec2. I personally think this is fine, since as far as I can tell we do no subsequent processing between when we check IUP and when we compile the final binary, but I want to minimize changes for the time being.

Another option would be to add another new type that is basically (Vec2, bool); a third option is to just return Vec<bool> with an entry for each input item, indicating if it can be interpolated.

I chose this last option, but after trying to integrate into fontc it feels bad; fontc passes ownership of the deltas to iup_delta_optimize, which means we need an additional clone of all the deltas in fontc in order to convert from our bools back to some useful representation.

Finally there's an additional headache; if we define a type in write-fonts we can't easily store that type in fontc contexts because write-fonts types do not implement the serde traits; this has come up before and so maybe is another argument for merging #392.

tl;dr: life is easiest of iup_delta_optimize returns GlyphDelta directly, and if this type implements Serialize and Deserialize?

On a happy note, if I use this patch in fontc than the gvar diff for our fonts drops dramatically, and it looks like the only remaining difference would be fixed by #614.

progress on googlefonts/fontc#422

[anthrotype]

I also think it would be fine if iup_delta_optimize API returns GlyphDeltas (which carry along the required flag), it's fine if they get rounded to integer as the caller doesn't need to process them further. And I don't like too much that we need to clone the deltas before calling iup_delta_optimize. Could we pass a reference or slice instead?

If we decide to return rounded GlyphDeltas, maybe the rounding method itself could be a parameter. Right now I think we don't match fonttools, which in the case of rounding of deltas it uses the built-in Python round() function instead of the usual OtRound (python built-in round() uses the Banker's method which is more statistically robust), but I am digressing and this belong to a separate issue (googlefonts/fontc#235)..

+1 for optional serde support

[cmyr]

I've added serde and updated this on top of it to return GlyphDeltas, and I think it's much cleaner overall.