Emit IR slows down the build considerably

```shell
$ hyperfine --warmup 3 --prepare 'rm -rf build/' 'target/release/fontc --source ../OswaldFont/sources/Oswald.glyphs --emit-ir=false'
Benchmark 1: target/release/fontc --source ../OswaldFont/sources/Oswald.glyphs --emit-ir=false
Time (mean ± σ): 149.9 ms ± 6.2 ms [User: 294.5 ms, System: 310.3 ms]
Range (min … max): 137.8 ms … 159.8 ms 18 runs

$ hyperfine --warmup 3 --prepare 'rm -rf build/' 'target/release/fontc --source ../OswaldFont/sources/Oswald.glyphs'
Benchmark 1: target/release/fontc --source ../OswaldFont/sources/Oswald.glyphs
Time (mean ± σ): 217.9 ms ± 11.6 ms [User: 409.5 ms, System: 1007.0 ms]
Range (min … max): 206.1 ms … 246.1 ms 10 runs
```

My guess, somewhat supported by a glance at flamegraph, is that this is because we block threads to emit IR. Tossing it over to another thread, on the rayon pool or otherwise, might be worth exploring as this would allow the processing threads to proceed.

Reproduce the issue with the two hyperfine commands shown in the report, comparing fontc with and without --emit-ir. Inspect the flamegraph and the emit-IR execution path to determine whether thread blocking causes the slowdown. Done means reducing the emit-IR performance cost while preserving the generated IR and existing behavior.

Written by the indexing model from the issue text.