Fix range reads on whole-file caches

[nvbkdw]

filecache and simplecache serve range reads from the local copy of a file rather than from the remote, and three things go wrong there:

1. Negative start/end are documented on AbstractFileSystem.cat_file to count backwards from the end of the file, and remote backends implement them as HTTP suffix ranges. _cat_file instead passed the value straight to f.seek(), so a negative offset raised OSError(EINVAL) after the whole file had already been downloaded.

2. The sync path reached AbstractFileSystem.cat_file, which needs f.size to resolve negative offsets. Both _open implementations return a bare io.BufferedReader, so it raised AttributeError. LocalFileOpener already attaches size to the raw file this way.

3. WholeFileCacheFileSystem._check_file returns a (detail, path) tuple while the SimpleCacheFileSystem override returns the path alone. _cat_file and _cat_ranges assumed the latter, so on filecache they raised TypeError/AttributeError for a cached path, and _cat_ranges never downloaded an uncached one (the miss sentinel is False, not None).

This surfaced with zarr v3 sharded arrays, whose chunk index lives at the end of each object and is read with a suffix range: every read through simplecache downloaded the full shard and then failed on the seek.