store: compute template digests while publishing

[CMGS]

Summary

• replace the pool-owned v1 full-tree pre-read and content_digest meta field with the canonical 16 MiB chunked v2 store contract
• make dir hash with a bounded per-call worker pool, store digest-<gen> beside immutable exports, and commit meta.json last
• hash the S3 source read used for upload, preserve multipart planning with explicit ContentLength, and return marker metadata without another object request
• propagate the committed digest through Promote -> Fetch -> Claim while keeping checkpoint/archive Publish records digest-free

The dir implementation is based on current main's generation-addressed layout from #67. Hash/upload/sidecar failures leave the previous meta, export, and digest visible; a fresh retry commits one complete replacement.

Validation

• make go-test
• make go-lint
• go test -race ./store/dir -run '^(TestFetchPinsGenerationAcrossStoreInstances|TestDigestWorkersStoreChunksByOffset|TestPublishDigestedFailureAndFreshRetry|TestSweepGenerationsPairsCurrentAndSupersededDigests)$' -count=20
• go test -race ./pool -run '^(TestPromoteThenClaimClonesFromTemplate|TestRepromoteContentDigestTracksExportBytes)$' -count=20
• Linux and Darwin ASL: store packages clean; the only pool output is the three pre-existing pool.go method-partition findings on unchanged main lines 584/605/624
• independent fixed-vector recomputation matched all published vectors, including 16 MiB - 1 / exact / + 1
• local MinIO RELEASE.2025-09-07T16-13-09Z: the 16 MiB + 1 real multipart upload matched the v2 vector; 64 MiB + 1 A/B (5 publishes/mode x 3) had overlapping ranges, with per-run medians plain 575/511/382 MB/s and digested 513/473/464 MB/s, so no fallback path was added
• M2 Max dir acceptance: a same-key claim during three 1 GiB re-promotes waited 88.8-151.5 ms

Production diff: +590/-214 lines. Test and test-support diff: +962/-301 lines.

Closes #64