the pyrefly-torch-stubs cause high cpu usage

[asukaminato0721]

pr open-spaced-repetition/srs-benchmark#315

[asukaminato0721]

oh, and ci tyck seems take too long.

[asukaminato0721]

cc @stroxler

[stroxler]

Thanks for the issue. A few notes:

• It looks like we're doing some late-binding work when applying DSL calls. This is a vibe-coding error I think (mine), it works but is not very efficient, the error probably would never have happened with human-written code.
• We probably could optimize the code path when gradual shape types come in so that there's less DSL overhead. This is probably less important than ^^, but would still help a lot in a large codebase with mostly-unannotated shapes
• I suspect the simple non-optimal factors above are the main drivers of poor performance today. That said, shape checking is inherently expensive, especially if our DSL evaluator is a tree-walk interpreter. In the much longer term we may want to optimize the interpreter itself, but we definitely need more confidence and stability first.

I'll try to get the lowest-hanging fruit in the next release (1.2) although I'm not 100% sure I'll have it ready in time.