Hash ExpnData without absolute positions [isolated: -3.8% incr-patched, serde -18%, 0 regressions]

[xmakro]

ExpnData hashed the absolute positions of its call site and def site. ExpnHash feeds the SyntaxContext hash of every macro-expanded span, so one inserted line at the top of a file changed the hash of every expansion below it and reddened all macro-expanded HIR in the crate.

The manual StableHash impl now hashes the two sites' hygiene context and skips their positions. Uniqueness comes from update_disambiguator, which numbers colliding expansions in expansion order. An edit that reorders identical expansions is a conservative miss.

Unsound as written, do not merge

The call-site position is itself a codegen input: walk_chain_collapsed collapses debuginfo to ExpnData::call_site, and #[track_caller] materializes Location from source_callsite(). Both read hygiene data untracked, so their only dep-graph coverage was the call-site position inside ExpnHash. With that removed, shifting a macro invocation leaves optimized_mir and downstream CGUs green, and reused object code keeps the old line numbers.

Repro (stage2 of this branch vs its base, same config):

// m.rs, insert a blank line above the invocation between sessions
macro_rules! mk { () => { pub fn foo<T>() { panic!("boom") } }; }
mk!();

// main.rs, not edited
mod m;
fn main() { m::foo::<u8>(); }

foo::<u8> instantiates into main's CGU, whose remaining inputs are untouched by the edit. After the incremental rebuild this branch panics at the stale m.rs:8:1 while a fresh build of the same source reports m.rs:9:1. The base compiler reports m.rs:9:1 both ways. Collapsed DWARF line info goes stale the same way. #[rustc_clean] confirms the flip: on the base compiler the edit dirties hir_owner(foo) and optimized_mir(foo), on this branch both stay clean.

Single-file tests and the incremental suite miss this because any edit changes the enclosing module's span, which reddens that module's own CGU and re-runs codegen, hiding the stale values. The cross-module instantiation above is the shape that escapes.

Since panic locations and debug line tables are observable outputs derived from the call-site position, codegen reuse must be invalidated when that position changes. A position-free ExpnHash cannot provide that on its own; the analysis-level reuse this PR buys is real, but it needs a separate mechanism to cover the codegen reads before it can land.