Better cache behaviour across embeddings versions

• Sampling npy points or regions from a local embeddings_dir no longer
  reads the whole directory firsts. (#384 @avsm @mdales)

• Downloading into an embeddings_dir that cannot be written now fails
  at once. The error names the number of missing tiles and the
  CLI command that would populate them. (#384 @avsm @mdales)

• Guard local caches against mixing dataset versions via a better
  cache key. zarr_store and GeoTesseraZarr now take cache_dir=
  and cache_max_size= to persist reads through zarr's CacheStore,
  keyed per store location so two datasets never share cached objects.
  NPY downloads into an embeddings_dir are also checked against a
  tessera_metadata.json sidecar and raise if the directory was
  populated from a different dataset. (@sadiqj @avsm)

• The zarr commands and the Zarr read API log progress through the standard
  logging module now rather than live progress bars, which previously
  corrupted multi-process output. Long reads always log through the
  geotessera.store logger at INFO (short runs stay silent); the
  progress= parameters on sample_points/read_region/iter_region/
  read_patch are ignored and will be removed in a future release.
  (@aneeshnaik @avsm)