Skip HIR hashing when the crate hash comes from encoded metadata [-0.97% non-incr full]

[xmakro]

Reworked on top of the upstream metadata-based crate hash work (upstream PR 154724, mirrored here as the upstream-154724-base branch). On that base the SVH is an XXH3-128 digest of the encoded metadata bytes plus a supplement fingerprint, and per-owner HIR hashing survives in non-incremental builds for exactly one job: compute_hir_hash sits in that supplement as blanket coverage for HIR-only inputs. This PR drops that job, which lets lowering skip HIR node and attribute hashing in non-incremental builds entirely.

Clean A/B against the base branch (full ThinLTO+jemalloc stage2, instructions:u, full matrix): -0.97% mean on non-incr full measurements over 164 cells, 100 wins / 2 regressions, 72 cells at or under -1%. syn check -2.45%, serde check -2.08%, hyper check -1.96%, cargo check -1.88%, externs check -6.34%. Incremental cells -0.07% mean (the change does not touch incremental hashing; the small residual win is the supplement no longer reducing every owner fingerprint). The only regression is the synthetic large-workspace at +0.43% debug full, where hundreds of tiny crates amplify the per-crate supplement walk.

What changes:

• needs_hir_hash no longer includes the plain needs_metadata case. Per-owner hashes are still computed for incremental (red/green fingerprinting of lower_to_hir), compilers built with debug assertions (query result fingerprinting), coverage, metrics, proc-macro crates, and -Zmetadata-crate-hash=no.
• The supplement seeded into the metadata hasher replaces the blanket HIR hash with:
  • for proc-macro crates: the full HIR hash, unchanged. Their metadata is a stub that does not describe the macro implementation, so the encoded bytes cannot stand in for the HIR.
  • for everything else: a hash of just the global_asm! owners (compute_global_asm_hash). Those are the one HIR-only input to codegen that leaves no trace in the encoded bytes — every should_encode_* predicate is off for DefKind::GlobalAsm, so only the def-kind discriminant is encoded regardless of the template — and the template can change without any source file changing, e.g. when a proc-macro reads the environment at expansion time. The owners are re-hashed from their nodes on demand (hash_owner_nodes_ungated), which is free in the common case of zero such items. This is what keeps tests/run-make/proc-macro-global-asm-changes-crate-hash passing without the blanket hash.
• The supplement composition depends only on the crate type and the flag, never on needs_hir_hash(): a compiler built with debug assertions, or an incremental build, produces the same SVH as a release compiler in a non-incremental build. The previous version of this PR keyed the scheme on needs_hir_hash() and had both warts.

What the SVH deliberately stops covering: HIR-only content that never reaches the encoded bytes and is not a global_asm! body — concretely, bodies of functions whose MIR is not encoded (not generic, not cross-crate-inlinable), in the case where those bodies come out of untracked expansion-time input rather than a source file. Anything that lives in a source file is still covered through the src_hash of files in the encoded source map; anything downstream compilation reads is in the bytes; tracked options are in the supplement's dep_tracking_hash.

tests/ui 21694 passed, tests/run-make 397 passed (including crate-hash-metadata-flag, the three proc-macro crate-hash tests, extern-flag-fun and emit-path-unhashed), tests/incremental 178 passed, 0 failed.