Suppress span-parent deps for embed-only queries [span-shift serde 25.0B -> 10.9B; -10% plain touch]

[xmakro]

Split from #88 (integrated stack and cumulative numbers there). Independent of the other pieces.

track_span_parent makes every query that inspects a parented span record a dependency on source_span(parent). That is the right conservative default, but for a small set of queries it only buys wasted re-execution: their results embed spans verbatim, derive nothing from the absolute positions, and every consumer rebases spans on reuse. For those, a position shift in the parent re-runs the query to produce a result that is identical except for the embedded positions.

thir_body is the amplifier that makes this expensive: it is no_hash, so it has no result fingerprint, so any red dependency makes it unconditionally red and drags mir_built and the rest of the MIR chain along.

This adds a TaskDepsRef::AllowIgnoringSpanParents variant, selected per DepKind in DepGraphData::with_task via suppresses_span_parent_deps(), for: thir_body, typeck_root, mir_built, check_unsafety, check_match, region_scope_tree.

Why this is sound:

• Suppression only affects span reads made directly inside the listed query's own task. Sub-queries record their own dependencies under their own tasks, so a const-eval query called from typeck still tracks source_span normally.
• Debuginfo stays correct because codegen tracks its own span reads. Verified empirically: a 3-line shift moves DWARF decoded line info by exactly 3.
• Diagnostics emitted by consumers of these results rebase spans on replay.

Perf: on the serde span-shift microbench (300 structs, one-line insertion at the top, instructions), this drops 24.95B to 10.88B on top of the two hashing fixes from #88, with no unstable flags involved. A plain-code control crate improves 10% on touch rebuilds (0.494B to 0.445B). On the collector (6 crates x check/debug/opt, jemalloc, geomean, on top of the hashing fixes): incr-patched:println goes from -5.39% to -7.13%, incr-full from +0.05% to -0.19%, no cell regressed at or above 0.25%. Not benchmarked fully isolated from the hashing fixes; needs its own perf run.

Validation (as part of the #88 stack): tests/incremental clean, full tests/ui clean (21309 tests), forced -Zincremental-verify-ich across 9 successive span shifts in both flag states with no panics, DWARF line tables verified against a shift, touch scenarios unregressed.

Siblings: #89, #90, #91, #93.

[xmakro]

Closing after re-measuring on top of #97 + #98. The re-execution this suppressed no longer happens: with the source_span anchors content-addressed (#97) and expansion contexts position-independent (#98), a span shift leaves those dependencies green, so there is nothing left to suppress. Rebased onto #98 and benchmarked (p94_spandep1_j vs p93_expn2_j): incr-patched -0.05%, i.e. the original effect is gone. A small residual remains from recording fewer edges (incr-full -0.25% geomean, 0 regressions), which would be better served by a dedicated dep-graph-size change than by this allowlist. The thir_body observation (no_hash, so a red dependency propagates with no early cutoff) is worth keeping in mind independently.