refactor: Introduce MemoHeader to reduce monomorphization

[MichaReiser]

Moves configuration-independent memo validation and cycle metadata into MemoHeader to avoid monomorphization.

This does not change the size of Memo because MemoHeader has no padding (it's always 32 bytes).

Release Binary size

Metric	Base	Branch	Difference
Stripped binary	23,189,920 B	22,430,096 B	−759,824 B / −3.28%
Unstripped binary	26,479,232 B	25,674,128 B	−805,104 B / −3.04%
Mach-O __text	16,110,760 B	15,403,432 B	−707,328 B / −4.39%

LLVM IR

Scope	Base	Branch	Difference
ty_python_semantic	12,484,695 lines	11,096,129 lines	−1,388,566 / −11.12%
Entire ty build	35,139,488 lines	33,380,017 lines	−1,759,471 / −5.01%

Testing: Passed local validation and GitHub CI.

[netlify]

✅ Deploy Preview for salsa-rs canceled.

Name	Link
🔨 Latest commit	9f5e215
🔍 Latest deploy log	https://app.netlify.com/projects/salsa-rs/deploys/6a3a389105243a0008075e40

[codspeed-hq]

Merging this PR will not alter performance

✅ 40 untouched benchmarks

---

_{Comparing MichaReiser:reduce-query-monomorphization (9f5e215) with master (a7d212d)}

[ibraheemdev]

A further opportunity here is to mark Memo as repr(C), allowing you to safely treat a *const Memo<C> as *const MemoHeader, and avoid the monomorphization entirely in some of the helper methods, such as provisional_status, which still rely on looking up the concrete memo type in this PR. I was able to get much larger wins on that branch by extracting some of our query execution logic into state machines as well, but I think this PR is a good first step, and it makes sense to get most of the boilerplate out of the way before making more targeted changes.