Add CacheDict.filepath and observe_access

[teonbrooks]

While working on our internal reporting tool, I've needed to patch these two features. They seem useful to upstream.

The first is to return the filepath from the cache so that it can be used to load parquets, etc.
The second is to build a graph of files used for the report. This allows us to export to minimally export a s3 bucket and it lets us do some treeshaking to remove cache files no longer needed for the reports.

• CacheDict.filepath(key): public accessor returning the absolute on-disk path backing a key (parquet/npy/pickle dumps, directory entries), where indexing returns the loaded value instead.
• observe_access(callback): context manager notified with the folder of every CacheDict constructed or accessed within its scope. Thread-safe; observer errors are caught and logged rather than propagated.

Adds unit tests for both.