Add concurrency control to LocalFileDescriptorStore #686

Currently, LocalFileDescriptorStore in server/app/backend/local_file.py has the same concurrency problems that #591 fixes for LocalFileIdentifiableStore (see #554):

• Multiple instances (e.g. one per uWSGI worker) can operate on the same directory. Each instance keeps its own _object_cache, which is not invalidated when another instance modifies the files.
• add() and commit() check for the file's existence before writing it, without holding a lock. A concurrent add() or discard() can create or delete the file in between, so writes are silently overwritten or a discarded descriptor is written back.
• add() and commit() write directly into the target file with open(..., "w"). A crash during the write leaves a truncated file, and a concurrent read can see a half-written one. LocalFileIdentifiableStore already avoids this by writing to a temp file and swapping it in with os.replace().

LocalFileDescriptorStore should get the same mechanism as LocalFileIdentifiableStore after #591:

• Acquire a DirectoryLock on the storage directory, including the disable_strict_locking opt-out for systems without fcntl
• Perform file and cache operations under a single _write_lock
• Write atomically via a temp file and os.replace()
• Add a close() method that releases the directory lock

Ideally, both stores share this logic, so that it does not have to be maintained twice.

This was deferred from the review of #591.