feat(bundler): per-provider tree-shake for vendor chunks

[lukekania]

Part of the v0.11.0 builder-parity milestone (issue 171 in the tracker stays open until the milestone branch lands).

Summary

• shake::collect_cross_chunk_used_names_per_provider returns a Vec<HashSet<String>> indexed by chunk index; for each chunk i it collects the names other chunks import from any module in i. Reuses the existing analyze_module parse + namespace-import expansion paths, adding a specifier → canonical-path fallback so bare npm specifiers ('@angular/core') attribute to the owning vendor chunk instead of falling through.
• bundle() derives that specifier-to-path map from the existing all_file_to_ns / specifier_to_namespace tables, then feeds externally_used_per_chunk[idx] into analyze_unused_exports for every chunk — the if is_main gate is gone. Vendor and lazy chunks now get the same cross-chunk consumption signal main has had since PR 131.
• npm_wrap::wrap_npm_module now accepts unused_exports and drops both the unused export const X = ... declarations and the matching __exports.X = ns.X re-export bridges, so shake decisions actually reach the emitted vendor chunk code (previously they died at the IIFE wrapper).

The old single-provider collect_cross_chunk_used_names and its expand_namespace_imports helper are removed — _per_provider subsumes them and computing the main-chunk set is just result[0].

Behaviour delta

On the test-ng-project fixture (production build), the largest vendor chunk shrinks from ~444 KB to ~416 KB. The gain on real-world @angular/core is modest because most of its exports are referenced by some module inside the package's own internal graph (so intra-chunk reachability keeps them alive even without entry-walk pinning); the bigger wins were left on the table by the old externally_used = None fallback when the package shape made the entry walk effectively reach everything. The unit and integration tests verify the mechanism end-to-end on a synthetic package with a true cross-chunk-only export.

Tests

• shake::tests::test_collect_cross_chunk_used_names_per_provider_multi_chunk — three-chunk graph (main/lazy/vendor), asserts per-provider sets are attributed correctly and unused names don't leak into the wrong chunk.
• shake::tests::test_collect_cross_chunk_used_names_per_provider_dotted_filename — port of the previous dotted-filename regression onto the new API.
• npm_wrap::tests::test_wrap_drops_unused_declaration and ..._reexport_bridge — unit-level proof that wrap_npm_module honours unused_exports.
• vendor_chunk_splitting_integration::vendor_chunk_drops_unreferenced_exports — full pipeline: a vendor chunk whose npm package re-exports two names from an impl.js declaration file emits only the consumed one in its bundled body.

Test plan

• cargo build --release -p ngc-rs
• cargo test -p ngc-bundler (97 tests pass)
• cargo clippy --workspace -- -D warnings
• End-to-end: production npx ng build against test-ng-project succeeds, vendor chunk shrinks measurably