…ested python_version/platform

Previously, when `configure_coverage_tool = True` was set but the bundled
`coverage.py` wheel set had no entry for the requested (python_version,
platform), `coverage_dep` returned None silently. The result was that
`bazel coverage` produced empty per-test lcov files for `py_test` targets
with no signal to the user that coverage was unconfigured.

Print a WARNING in that path so the misconfiguration is visible. Preserve
the existing silent return for the windows branch, which is intentionally
quiet because the upstream coverage wrapper does not support windows.

Per review on bazel-contrib#3766, replace the raw `print(...)` warning with
`logger.warn(...)` from `repo_utils.logger`, with the logger threaded
in from the caller.

- `coverage_dep` takes an optional `logger` parameter and falls back to
  a default-constructed logger if none is supplied.
- `python_register_toolchains` accepts a private `_internal_module_ctx`
  kwarg (mirroring the existing `_internal_bzlmod_toolchain_call`
  pattern). When invoked from the bzlmod path, it builds the logger
  with the real `module_ctx` so module-root filtering applies (see
  bazel-contrib#3760). For the
  WORKSPACE/macro path it constructs a minimal stand-in struct, which
  is all the logger needs.
- `python.bzl` passes `module_ctx` through.

Adds tests/coverage_deps/ with two cases (unsupported version warns,
windows stays silent) using the captured-printer pattern already
established in tests/pypi/hub_builder/hub_builder_tests.bzl.

Add a comment explaining that the supported-wheel path of coverage_dep
calls maybe(http_archive, ...) which calls native.existing_rule(),
which is only valid during BUILD/macro/finalizer evaluation, not during
rule analysis where rules_testing analysis tests run. The path is
covered end-to-end by real bazel coverage runs.