Retry S3 chunk validation failures and preserve the underlying error

[folbricht]

Fixes #334.

Problem

desync untar against an S3-compatible store (Cloudflare R2 in the report) intermittently failed with:

chunk id <ID> does not match its hash 0000000000000000000000000000000000000000000000000000000000000000: unexpected EOF

The object in the store is fine. The real cause is a transient short read of the chunk body somewhere between R2 and minio-go/the HTTP transport. When that happens, io.ReadAll in S3Store.GetChunk returns truncated bytes, NewChunkFromStorage → Chunk.ID() → Chunk.Data() fails to zstd-decode the partial stream, but Chunk.ID() discards that error and returns the zero ChunkID{} — so it's reported as a hash mismatch against 0000…, hiding the real error. And GetChunk's --error-retry loop only covered GetObject and io.ReadAll, not NewChunkFromStorage, so the transient failure was fatal.

Changes

• errors.go — ChunkInvalid gets an optional Err (underlying cause) and an Unwrap(); Error() reports the cause when present instead of the bogus zero-hash message.
• chunk.go — NewChunkFromStorage / NewChunkWithID call c.Data() explicitly and return ChunkInvalid{ID: id, Err: err} on a decode failure, rather than letting ID() swallow it. Still ChunkInvalid, so RepairableCache and verify --repair keep self-healing such chunks.
• s3.go — S3Store.GetChunk retries NewChunkFromStorage validation/decompression failures under the existing --error-retry policy (same backoff as the other retry blocks), logging each retry, and closes the object reader before looping so deferred closes don't stack.
• Tests — new s3ErrCorruptBody mode in the fake S3 server (well-formed HTTP response with a truncated body — the issue's scenario) plus corrupt_body_fail / corrupt_body_recover subtests, and a new chunk_test.go covering the ChunkInvalid.Err paths.

go build ./..., go vet ./..., gofmt -l ., and go test ./... all pass.

Out of scope

gcs.go and sftp.go GetChunk have no --error-retry loop at all, and remotehttp.go has one but doesn't cover NewChunkFromStorage. They all benefit from the improved error message here, but adding retry loops there is a separate change — happy to include it if wanted.