perf: dep_graph: stop collecting edges in the green marking walk [-1.70% incr-unchanged]

[xmakro]

Based on the parent branch perf/carry-unchanged-nodes, where dep-graph indices hold still across sessions and an unchanged node's record is re-emitted from the previous file.

Marking a node green walks its previous dependencies and, on the way, copies each dependency's current index into a scratch buffer. That buffer exists for one reason: to hand the encoder an edge list for the node it is about to write.

The encoder no longer reads it. A promoted record comes out of the previous file as it stands, edges included. The one remaining reader is the debugging graph behind -Zquery-dep-graph, and it can take the edges off the previous graph instead: every target of a promoted node is green, and a green node keeps its previous index, so the current edge list is the previous one unchanged.

So the walk stops collecting. This removes a store per edge from the hottest incremental path, along with the per-node edge frame and its drop, the per-worker buffer behind it, and the edge list threaded through promotion.

Local A/B against the parent branch, six primary crates, instructions:u, jemalloc preloaded, both sides built from an identical bootstrap.toml:

scenario	delta
incr-unchanged	-1.70%
incr-patched	-1.06%
full / incr-full	+0.03% / -0.01% (control)

43 of 81 cells improved by at least 0.25%, 0 regressed. Best cells: cargo debug incr-unchanged -2.22%, syn debug incr-unchanged -1.97%, regex-automata debug incr-unchanged -1.91%.

Validation, with debug assertions and overflow checks armed: tests/incremental 178 passed and 0 failed, of which the 93 tests that pass -Zquery-dep-graph exercise the reconstructed edge lists. Sixteen generations of churn and deletion stress in each of four modes (default, -Zincremental-verify-ich, -Zthreads=8, and both together) came back clean, including the warning-replay oracle, and a multi-generation -Zquery-dep-graph build is clean as well.