fix(pypi): skip a pip.parse version with no interpreter for the host

[zakinko]

On Windows ARM64, evaluating the pip extension fails as soon as any
module in the graph calls pip.parse for Python 3.9 or 3.10:

Unable to find interpreter for pip hub 'grpc_python_dependencies' for python_version=3.9: ...
Expected to find python_3_9_host among registered versions:
  python_3_11_host
  python_3_12_host
  python_3_13_host

There is no CPython 3.9 or 3.10 build for aarch64-pc-windows-msvc in the
runtime manifest, so those toolchains have no host interpreter there.

_pip_parse already handles a close relative of this: a version missing
from minor_mapping is skipped with an info message, and the hub's
select fails later only if that version is actually used. This treats a
version with no host interpreter the same way, right after that check.
The python_X_Y_host name is built in one small helper, so the new check
and _detect_interpreter look at the same thing.

To check it, I ran bazel query @bazel_pip_dev_deps//:all in a plain
checkout of Bazel master (rules_python 1.9.2, grpc 1.76.0) with and
without this change back-ported to 1.9.2, on windows-11-arm and
ubuntu-latest:

                 without   with
windows-arm64    rc=7      rc=0    3.9 and 3.10 skipped for grpc, protoc-gen-validate
                                   and rules_fuzzing
linux-x86_64     rc=0      rc=0    nothing new skipped

https://github.com/zakinko/NetBSD-i386/actions/runs/36248007899

With RULES_PYTHON_REPO_DEBUG_VERBOSITY=INFO the Windows run prints the
new message once for each of the six hub and version pairs, next to the
existing "no registered toolchain" one for rules_fuzzing's 3.8.

The new test in hub_builder_tests.bzl fails without the change with the
same error, and //tests/pypi/... passes with it (254 tests, on macOS).

[rickeylev]

Thanks for the fix!