* filer: do not 404 a TUS session on a transient chunk-load failure
readTusSessionInfo already proved the session exists before
loadTusSessionChunks is called, so a failure there is a read failure,
not evidence the session is gone: a volume-server timeout or a
canceled request context surfaces through ListDirectoryEntries the
same way a missing session would.
Every such error was mapped to writeTusSessionNotFound, answering 404
to HEAD/PATCH and 204 to DELETE. A spec-compliant TUS client trusts
that and discards the session, orphaning every chunk it had committed
until the 24h expiry sweep, or forever if it never issues a DELETE.
Only an error matching filer_pb.ErrNotFound is now reported as not
found; anything else answers 500 so the client retries against the
same session instead of abandoning it.
* test: cover a TUS session's transient chunk-load failure
Adds a listErr hook to the in-memory test store, alongside the
existing commitErr/deleteErr, to simulate a store or RPC failure from
ListDirectoryEntries.
HEAD, PATCH and DELETE against a live session all answer with a
server error instead of a not-found status when the chunk listing
fails transiently, and the session is left on disk untouched. A
listing failure that genuinely means not found, filer_pb.ErrNotFound,
still answers 404 (204 for DELETE).
