Process-global manifest cache is keyed only on manifest_path

Start in pyiceberg/manifest.py by reading _ManifestCache.get_or_cache and the read_manifest_list entry point. Reproduce the collision with two manifest lists that use the same manifest_path but provide different metadata. Done means cache behavior no longer allows one manifest list's ManifestFile metadata to affect another table's scan pruning, with regression coverage for the collision.

pyiceberg/manifest.py holds a module-level cache:

_manifest_cache = _ManifestCache()

Its docstring describes it accurately: "Process-wide ManifestFile cache keyed by manifest_path." The key is the path string alone — no catalog, table, session, or tenant forms part of it.

def get_or_cache(self, manifest_file: ManifestFile) -> ManifestFile:
    ...
    manifest_path = manifest_file.manifest_path
    if manifest_path in self._cache:
        return self._cache[manifest_path]
    self._cache[manifest_path] = manifest_file
    return manifest_file

On a hit it returns the cached object and discards the one just read. Since read_manifest_list constructs ManifestFile objects from the entries of a manifest list rather than by reading each manifest file, the cached object reflects whatever the first manifest list to name that path declared — including its partition summaries, counts and sequence numbers.

Two tables in the same process that reference the same manifest_path therefore share one ManifestFile, and the first read wins. Because those fields drive scan pruning, a stale or mismatched entry changes which files a scan considers.

Issue investigation generated via claude, reviewed by Sung, Kevin, Fokko.