Return type op fast path results without the full fulfillment machinery [-0.83% avg]

[xmakro]

Query type ops check their fast paths (trivial sizedness proofs, alias free normalizations) inside the closure passed to scrape_region_constraints. Every fast path hit therefore still paid for an inference snapshot, an ObligationCtxt with a boxed fulfillment engine, RefCell take and reset cycles for region constraints, and the region constraint scraping itself, all over empty data. MIR type check issues several type ops per statement, normalizing the place and rvalue types and proving sizedness of the place type, and the overwhelming majority of these hit a fast path.

Fast path results are final. They register no obligations and produce no region constraints, so the fast path is now checked once at the top of fully_perform and the result returned directly, skipping all of that machinery. The two inner fast path checks become dead and are removed. resolve_vars_if_possible is kept on the returned value because the normalize fast path can return values that still contain inference variables, matching what scrape_region_constraints did before.

Local measurement: from scratch ThinLTO stage2 with jemalloc, instructions:u, full rustc-perf suite, both sides built from scratch.

• primary benchmarks: mean -0.83% across all cells; check profile mean -1.34%, debug -0.71%, opt -0.45%
• check full on large primaries: image -3.99%, ripgrep -3.84%, cargo -3.81%, syn -3.48%, cranelift-codegen -3.42%, hyper -3.13%, diesel -2.16%
• statement heavy stress benchmarks: ucd -7.1%, unicode-normalization -6.5%, tuple-stress -6.1%, deep-vector -5.4%, coercions -5.3%
• 256 cells improved by at least 0.25%, 1 regressed (a single known noisy incr-patched cell)

tests/ui/borrowck, tests/ui/nll, tests/ui/sized-hierarchy, tests/ui/associated-types and tests/ui/higher-ranked pass (1752 tests, 0 failures). x check compiler is clean.

---

Update: review found that skipping scrape_region_constraints also skipped its role as the unconditional drain point for region state pending on the InferCtxt. The old-solver hidden type equation in insert_hidden_type registers region constraints eagerly and could leave partial state pending when it fails partway (the ? in relate_opaques bails before register_goals scrapes). Reworked in two commits:

1. relate_opaques now rolls back partial hidden type equations via commit_if_ok, making "no region state pending between type ops" a real invariant.
2. The hoisted fast path debug-asserts that invariant (plus not-in-snapshot) via a new InferCtxt::has_pending_region_state helper instead of relying on it silently. Release behavior and the perf numbers above are unchanged.

Note: the next solver's solver_region_constraint_storage (-Zassumptions-on-binders) is cross-op by design and intentionally outside both the scrape and the assertion.

Full tests/ui (21304 tests) passes on a stage1 build with debug-assertions = true, exercising the new assertions.