fix: accept multipath in the site cache functions

[darrenhuai]

multipath changes four site directories: site_data_dir and site_config_dir on Unix and macOS, site_applications_dir on both, and site_cache_dir on macOS when the interpreter comes from Homebrew. The module-level functions expose the argument for the first three, but site_cache_dir() and site_cache_path() take no multipath at all, so passing it is a TypeError and the Homebrew cache prefix is unreachable from the function API:

>>> platformdirs.site_cache_dir("SuperApp", multipath=True)
TypeError: site_cache_dir() got an unexpected keyword argument 'multipath'
>>> platformdirs.PlatformDirs("SuperApp", multipath=True).site_cache_dir
'/opt/homebrew/var/cache/SuperApp:/Library/Caches/SuperApp'

docs/explanation.rst already documents the behaviour those functions cannot produce ("site_data_dir and site_cache_dir include the Homebrew prefix as an additional path when multipath=True").

Adding it keyword-only keeps opinion and ensure_exists where they are for positional callers, and matches how the applications functions took their new options in #534/#535.

On site_cache_path it is a no-op, since the path variants return the first entry either way. I kept it for the same reason site_data_path and site_config_path carry it — test_function_interface_is_in_sync requires the dir and path signatures to match.

The new test parametrizes over the whole multipath family rather than just the cache case, so the next site directory that grows a multipath variant has to be reachable from its function too. Also corrected the multipath docstring in api.py, which still claimed only site_data_dir and site_config_dir were affected.

Impact is narrow, and worth stating plainly: it only bites a macOS user on Homebrew Python who wants both cache roots and is using the function API rather than PlatformDirs. It is an API-completeness gap, not a crash anyone has reported.

Verified locally on Windows: full suite 1249 passed / 94 skipped, ruff check src tests, ruff format --check, ty check --error-on-warning . all clean, and the guard test fails both when the change is reverted and when the signature is kept but the argument is not forwarded. I did not build the docs (no sphinx in my venv); the docstring edit is a plain literal.