feat: Store.head content-version probe + parallel chunk decode (v0.7.2)

[guidocerqueira]

Two changes, two commits.

1. Store.head for content-version probing (4edc6e3)

Adds an optional Store.head(key) + S3Store.head returning { etag, lastModified, size } | null (null for an absent key; other failures throw, mirroring has()'s retry + 404 handling) and a StoreHead type.

Why: the ZarrDatasetRegistry keys every cache tier (handle/L1, metadata + coordinate Redis caches, on-disk chunk cache) by the open() id and assumes a dataset is immutable per id. When an upstream pipeline overwrites a dataset in place, every tier kept serving stale state — in nautilus-api this surfaced as a collapsed/constant forecast time axis that survived a Redis flush and only cleared on a pod restart. head() exposes the ETag as a cheap content-version probe so a consumer can fold it into the cache key (id = \${s3Path}@${etag}``) and invalidate all tiers at once on re-ingestion.

• src/store/store.ts — optional head? + StoreHead.
• src/store/s3.ts — S3Store.head() (retry + AbortSignal timeout; null on 404).
• tests/integration/s3.test.ts — head ok + 404→null (local fake S3).

2. Parallel chunk decode via worker pool (d759100)

Off-thread Blosc/codec decoding through a worker pool (decode-pool + decode-worker), wired into the chunk loader so the models x variables x chunks fan-out no longer serializes decode on the main thread. Adds a benchmark, unit coverage, and updates the CJS/ESM interop tests + README for the new export.

Release

Bumps to 0.7.2 + CHANGELOG (head feature). After merge, tag v0.7.2 to publish.

Test

typecheck clean; full suite 344 passing / 17 skipped; lint + format clean.