Make -Zcache-proc-macros survive span shifts [touch -21%; shift edits: zero hits -> -10% vs uncached]

[xmakro]

Split from #88 (integrated stack and cumulative numbers there). Based on the attr-arg span parenting branch, which provides the TokenStream::map_spans utility; everything else is flag-gated and independent of the other siblings.

-Zcache-proc-macros currently gets zero cache hits on any edit that shifts positions: the query key is (LocalExpnId, &TokenStream), and both components are position-fragile. Expansion indices shift whenever an edit changes how many expansions precede a derive (measured: a one-line comment insertion at the top of the file misses 598 of 600 cached expansions), and the token spans in the key shift with the edit. On a 300-struct serde crate the flag was a net regression on such edits: 42.0B instructions with the flag vs 41.3B without.

Three changes, all in the flag path:

• Span-agnostic keying. A specialized DepNodeKey impl fingerprints the key as (parent module DefPathHash, macro DefPathHash, tokens hashed with span hashing disabled). The triple is collision-free for legal code: identical tokens include the item name, and duplicate names in one module are rejected, so no two live expansions can produce the same fingerprint. LocalExpnId deliberately does not enter the hash. ExpnId::stable_hash asserts default hashing controls, so the pieces are hashed manually rather than hashing the whole key under hash_spans = false.
• no_hash on the query. A span-only difference in the output tokens is deliberately treated as no change; that is the contract that makes reuse across shifts possible. Without it, incremental_verify_ich (which re-hashes roughly 1 in 32 cache loads even when the flag is off) panics with unstable fingerprints when the reused output's spans have shifted.
• Anchored respan of cached output. Cached token spans used to round-trip through line/col decode against the current file, so the reconstructed offsets drifted every session. Output tokens are now respanned to a file-start anchor with root context, keeping each token's own syntax context. The anchor choice matters: respanning to the call site makes source_span of every generated def move with edits, which re-runs typeck and MIR building for all generated bodies; the file-start anchor keeps them position-stable.

Perf (300-struct serde crate, instructions): touch rebuilds 8.25B to 6.50B with the flag (-21%; the serde_derive dylib run alone is 16% of rustc time). Span-shift edits go from zero hits (+1.9% vs no flag) to hitting: -10% vs no flag measured standalone on the original base; on top of the full #88 stack the same edit is 9.05B with the flag vs 10.9B without. Adds an incremental test that a span shift reuses the cached expansion (loaded_from_disk).

Tradeoff, flag-only: errors reported inside a cached expansion lose their exact position; help notes and expansion notes stay correct and line accurate because ExpnData is up to date. Flag-off diagnostics are unchanged.

Validation (as part of the #88 stack): tests/incremental clean, full tests/ui clean (21309 tests), forced -Zincremental-verify-ich across successive span shifts with the flag on with no panics, diagnostic fidelity checked for an error inside a cached expansion, serde_json output of a cached-then-shifted build round-trips correctly, mid-file derive insertion correctly misses instead of reusing stale output.

Depends on #89 (provides map_spans; this PR is based on that branch). Siblings: #90, #91, #92.