Replace analysis-time depset.to_list() loop with Args.add_all(map_each=...)
and ctx.actions.write(Args), deferring depset expansion to execution time.

The old code built the input file list via string concatenation in a loop,
which is O(n^2) in the total string length. For large py_wheel targets this
dominated analysis time (~18s for ~18k transitive files in a production
monorepo).

Tested under Bazel 8.5.1 and 9.0.0.

Add tests/py_wheel_performance/ with:
- A correctness test verifying wheel contents (100 deps)
- A scaling regression test (tagged manual) asserting analysis time
  grows linearly, not quadratically, with dep count (5k vs 10k deps)

Scaling test results (5k vs 10k deps, best of 3):

  Before: 2825 ms / 8616 ms = 3.05x (quadratic)
  After:   502 ms /  776 ms = 1.55x (linear)