[fsspec-aware #479] zarr: make save_to_disk write path fsspec-aware (write, priority 1)

[xroynard]

Sub-task of #479. Write side, zarr backend — easiest, do first.

Goal

Let the zarr writer target any fsspec URL (file://, s3://, gs://, http://) instead of being pinned to a local store, and write per sample so peak local disk drops to ~1 sample.

Current state (verified on main)

• zarr/writer.py iterates write_sample per sample already, but the store is a concrete zarr.storage.LocalStore.
• The per-sample loop is the right hook; only the store selection is hardcoded.

Scope

• Derive the store from the target path / storage_options (fsspec-aware) instead of pinning LocalStore.
• Keep the existing per-sample write loop; writes go directly to the target as produced.
• Gate optional protocol deps (s3fs/gcsfs/…) behind a [remote] extra with a clear error when the handler is missing (per fsspec-aware save_to_disk: stream samples to the target during writing (avoid full local materialization before remote push) #479 open questions).

Notes / traps

• fsspec handle management under num_proc > 1 (independent, process-safe handles per worker) — see fsspec-aware save_to_disk: stream samples to the target during writing (avoid full local materialization before remote push) #479.
• zarr writes needing a seekable handle may need a small per-sample local temp buffer, uploaded then discarded (still bounded peak disk).

Tests

• Round-trip on memory:// (write then read) so it is CI-testable without network.

Priority

★ (1/5) — smallest, highest value, unblocks the incremental-write use case from #478/#479.