[feaLib] SingleSub compilation not always optimal

[cmyr]

This is a slightly vague issue title because I do not know the full extent of the problem, but I do have one specific example, which is demonstrated by the spec5f_ii_4.fea test case in feaLib.

This case involves some inline substitution rules. feaLib combines all of these rules into one SingleSub format 2 subtable, but it is more efficient to put most of the rules in a format 1 subtable; this optimization would would reduce the generated binary from 106 bytes to 52.

I don't have strong feelings about whether or not this is worth fixing; I'm just trying to debug a few of the remaining places where fea-rs and feaLib are generating different ttx, and this is one of those.

[anthrotype]

@cmyr can you share the code in fea-rs which applies this optimization? Maybe we port this over to otlLib

[anthrotype]

I think it's this one:
https://github.com/cmyr/fea-rs/blob/f4477919dc9658d9609464df7719d75e53832912/fea-rs/src/compile/lookups/gsub.rs#L62-L149

[cmyr]

Yep, that's it. At a high level:

• we collect all rules into a map of target: (replacement, delta)
• we then sort them into two groups, based on whether they are eligible for format 1 or not
• we then go through all the candidate format 1 groups and see if the savings justify the cost, moving items to format 2 where it is beneficial.

There are some possible improvements to this, but I believe any remaining gains are going to be very marginal.