linker performance tuning, round 1

[mighdoll]

Check to see if there's any obvious performance gotchas in the linker.

[stefnotch]

I could update the bench to have a mode where the entire linker is executed. Would that do the trick?

[mighdoll]

The thought here was for a quick perf review in the M1 timeframe. @stefnotch has already improved the perf quite a bit at the parsing layer, but the other parts of the linker are unexamined perf wise.

I was thinking roughly:
. collect 1-2 realistic larger shader projects as test cases (e.g. perhaps a bevy subset and one other project)
. run with performance sampling (e.g. via bench --profile), (extended to link() as you suggest!). node performance should be close enough to the browser for this test.
. also test a minified production build to see that it tracks proportionally

and fix any silly perf wins easily available.

Marking as M1+.