Bugfix: Scope macro digest cache to each hash invocation

[dgollahon-plaid]

The query service can omit targets affected by a macro-only .bzl change. It hashes the base and head revisions in one process, while RuleHasher is a singleton. PR #436 added a .bzl digest cache keyed only by workspace-relative path, so hashing the head revision could reuse the base revision’s digest and produce identical target hashes.

Move revision-derived .bzl cache state into a HashInvocationContext created for each BuildGraphHasher invocation. The context is shared across parallel target hashing, which preserves same-run digest deduplication, but it cannot leak across base and head checkouts. This also avoids an unbounded process-lifetime cache and remains safe if hash generation becomes concurrent.

The regression was introduced by 700780a in #436. The same two-revision query-service reproduction:

• passes at its parent, 07b6d54, where //:generated is impacted;
• fails at 700780a, where the service returns no impacted targets;
• passes with this change.

The regression coverage also verifies that .bzl digests are recalculated between hash invocations and still deduplicated within one parallel invocation.

[dgollahon-plaid]

This was discovered by A/B testing old/new bazel-diff on Plaid's codebase for a representative set of changes. The new version of bazel-diff omitted changes that should have been triggered. I had Codex produce a regression test and a fix. I am not personally familiar with this codebase so happy to implement any feedback or alter approach. You can also view this PR as a bug report instead if you prefer to implement something different.

Thanks for a valuable tool!