fix(uv): deduplicate whl_install wheel file labels

[jimmyt857]

Summary

uv.lock can contain multiple wheel entries that resolve to the same fetched wheel repository label. One way this happens is when duplicate package records have wheel URLs with different basenames but the same hash, so rules_py fetches them through the same http_file repository.

Before this change, the uv extension forwarded those duplicate labels directly to the whl_install.whl_files label_list attr. Bazel rejects duplicate labels during repository rule instantiation, so the install fails before wheel metadata can be extracted.

This change deduplicates the non-empty wheel file labels while preserving order, then changes whl_install metadata extraction to map the deduplicated labels back to the serialized prebuild targets by label string instead of relying on positional pairing. That keeps the label_list valid without losing top-level metadata for selectable wheels.

Regression Test

Adds e2e/cases/uv-duplicate-whl-labels, whose lockfile contains two six==1.16.0 wheel entries with the same hash and different wheel basenames. Without the fix, the test fails during repository rule setup with a duplicate whl_files label error.

Testing

• bazel run //:gazelle
• bazel test //uv/private/extension:all //uv/private/whl_install:all
• (cd e2e && bazel test //cases/uv-duplicate-whl-labels:test_import //cases/uv-dup-lock-records:all)
• pre-commit run buildifier --files e2e/MODULE.bazel e2e/cases/uv-duplicate-whl-labels/BUILD.bazel e2e/cases/uv-duplicate-whl-labels/setup.MODULE.bazel uv/private/extension/defs.bzl uv/private/whl_install/repository.bzl

Note: pre-commit run buildifier-lint --files ... still reports existing warnings in touched upstream files, so I kept this patch focused on the regression rather than sweeping unrelated lint cleanup.

[github-actions]

py_binary startup benchmark

Version	Mean (ms)	Median (ms)	± stddev	vs BCR	vs main	Build (s)
BCR 1.11.7 (baseline)	193.037	190.524	±9.350	—	—	40.85
HEAD main	60.817	60.708	±0.591	-68.5%	—	9.59
This PR	61.749	60.649	±4.701	-68.0%	+1.5%	6.93

Measured with hyperfine --warmup 5 --runs 50 on Linux
Gate: PR vs HEAD main (threshold: 10%). BCR is shown only as a historical baseline.
Build time: cold bazel build //:bench with isolated output base, no disk cache.

sys.path quality

Version	sys.path entries	distinct site-packages roots	duplicate realpaths
BCR 1.11.7 (baseline)	6	1	0
HEAD main	7	2	0
This PR	7	2	0

sys.path quality measured by bench_syspath inside the assembled venv. Duplicate realpaths indicate symlink redundancy; many distinct site-packages roots suggest an inefficient venv layout.

✅ No regression detected (PR is +1.5% vs HEAD main)