Fix GCS not-found detection to unwrap ErrObjectNotExist (#358)

[folbricht]

Problem

desync make fails when it encounters a chunk that doesn't yet exist in a Google Cloud Storage bucket, aborting with:

Error: storage: object doesn't exist: googleapi: Error 404: No such object

The missing object is expected — it's a new chunk awaiting upload — and should be treated as a cache miss, not a fatal error.

Root cause

gcs.go detected "object not found" using direct equality (err == storage.ErrObjectNotExist) in three places (GetChunk after NewReader and after io.ReadAll, and HasChunk after Attrs).

The dependency bump in a452c77 (cloud.google.com/go/storage v1.30.1 → v1.60.0) changed the library to wrap the sentinel rather than return it bare. formatObjectErr now does:

return fmt.Errorf("%w: %w", ErrObjectNotExist, err)

so the returned error is no longer == storage.ErrObjectNotExist. Both Attrs and NewReader route through formatObjectErr, producing exactly the "storage: object doesn't exist: googleapi: Error 404: No such object" message above.

In ChunkStorage.StoreChunk, the dedup check if hasChunk, err := s.ws.HasChunk(...); err != nil || hasChunk then propagates this misclassified error and aborts the whole make. The same flaw also turns legitimate read misses in GetChunk into hard errors instead of ChunkMissing, breaking cache/failover self-repair. (S3 is unaffected; it uses status-code-based detection.)

Fix

Use errors.Is, which unwraps the chain, in all three locations. No import change is needed — github.com/pkg/errors (already imported) passes Is straight through to the stdlib.

Fixes #358