Hash expansion call sites position-independently so macro-expanded code survives shift edits

[xmakro]

Stacked on #99 and #97 (the fused def_anchor stack), which this depends on. Reworked from the earlier revision that sat on the closed #91's channel; the expn_call_site_position position-fold query from that revision is replaced by expn_anchor, the expansion-side twin of def_anchor.

ExpnData is the remaining position carrier after #97: its call_site and def_site spans hashed absolutely, so editing text above a macro invocation re-fingerprinted the expansion and everything expanded from it, even though relative spans and the content-addressed source_span anchor were already position-free. This PR hashes both spans position-independently (stable_hash_span and stable_hash_anchor_span become one implementation differing only in whether the absolute arm covers the offset; parented spans were already relative). Identity between textually identical invocations rests on the existing disambiguator mechanism, plus one lowering fix: mark_span_with_reason now parents desugaring call sites the way rustc_expand already parents macro invocation spans, so same-kind desugarings in different definitions are distinct by parent rather than by session-order disambiguation. That also turns a previously order-dependent typeck_root invalidation in tests/incremental/hashes/function_interfaces.rs deterministic; the test's annotations are updated.

The anchor

With positions out of the expansion fingerprint, a renderer that lands on a parentless collapsed call site (debuginfo for macro-generated definitions under #[collapse_debuginfo], #[track_caller] whose expansion cause is an item-level invocation) needs an explicit dependency, exactly as definition-extent renderings need def_anchor. That dependency is expn_anchor(ExpnHash): the same content as def_anchor, the file id plus SourceFile::line_extent_hash over the call-site span, so it covers the rendered line index, column and extent line structure of the call site and nothing else. A byte-shifting edit above the invocation that moves no line break stays green; an edit that changes any rendered value goes red. The previous revision paired an opaque position fold with a separate line-table dependency at each site; here it is one node with the precision of the rendered values, the same fusion #99 applies on the definition side.

The dep-node key is the ExpnHash itself: it packs losslessly into the key fingerprint, so this PR reintroduces the small KeyFingerprintStyle::SelfHash recovery style (three hunks) that the bucketed design had used, now scoped to this one key type. Forcing a node for a foreign expansion resolves it through the defining crate's metadata; expn_hash_to_expn_id becomes fallible so a vanished crate or expansion keeps the zero sentinel and re-executes dependents instead of ICEing, while the on-disk-cache decode path (whose references can only have been green if the expansion still exists) unwraps loudly.

Wiring

TyCtxt::walk_chain_collapsed_tracked replaces the untracked hygiene walk wherever the collapsed result feeds a cached artifact. When the collapse lands on a parentless call site it records expn_anchor and registers its extent for #99's line-observation hook; otherwise the caller anchors through the span's parent or the instance as before. A collapsed rendering that escapes both anchors trips the hook's conservative fallback and costs reuse rather than staleness.

Expansion anchors are recorded only where no definition-keyed anchor already covers the position. An ExpnHash is fragile across edits: inserting any macro invocation (the benchmark suite's println! patch, for instance) shifts the disambiguators of later identical invocations, changing their hashes, and a dep node whose key cannot be recovered can never be marked green, so every dependent goes red. Definition-keyed anchors are stable under the same edit, so they take priority and the ExpnHash-keyed nodes are reserved for positions nothing else covers (this took ripgrep's patched rebuild from seven modules re-codegened to one, against the baseline compiler's five). Bang invocations hash their call-site extent; derive and attribute invocations, whose recorded call sites are just the path, stay file-granular. Wired at the cg_ssa debuginfo funnels, cg_clif's get_span_loc, cg_llvm's file_metadata_from_def_id, and span_as_caller_location via the new expansion_cause_with_expn.

tests/run-make/incr-macro-static-decl-line pins the edge: a #[collapse_debuginfo] macro generates a static, and the two byte-length-identical versions swap the invocation line with a padding line, so nothing re-fingerprints and only the expn_anchor dependency refreshes the static's DW_AT_decl_line.

Performance

Same-day A/B on the setup from #99 (instructions:u, jemalloc, 81 cells). Marginal effect of this commit against the stack head (#99 + #97):

scenario	marginal vs stack
full	+0.05%
incr-full	+0.04%
incr-unchanged	-0.86%
incr-patched	-3.31%
all cells	-1.28%

The largest single contributor in the stack. The patched win is where expansion-heavy code lives: serde incr-patched improves about -18% across profiles, and cargo opt incr-unchanged improves -18.6% (this commit's anchors give the caller-location and collapsed-debuginfo classes their precise invalidation channel, which #99 standalone leaves to conservative re-execution). The cost side is a broad set of small incr-unchanged cells at +0.28% to +0.85%, the marking cost of the expn_anchor nodes; the coverage-first recording rule above keeps their count down.

Cumulative for the full reworked stack (#99 + #97 + #98) against the merge-base (fresh base run), 81 cells: full -0.15%, incr-full +0.07%, incr-unchanged -0.47%, incr-patched -4.03%, all cells -1.48%, with 47 cells improved by at least 0.25% and 8 regressed; every check-profile cell improves (serde incr-patched -19.0% to -20.7%). The largest residual is cargo debug incr-unchanged +9.4%, with the rest at or below +1%. See #99 for the trust-tiered anchor design and the staleness bugs the fail-red rework surfaced.