Copy S3 objects server-side

[sduchesneau]

S3Store.CopyObject used to download the object and upload it back, moving every byte through the client and, on a compressed store, decompressing and recompressing it on the way. It now asks the service to do the copy: a single CopyObject call for objects up to 5 GiB, the largest S3 accepts in one call, and a multipart copy of 1 GiB parts above that, aborted if any part fails. GSStore.CopyObject already worked this way.

Verified against a local MinIO (docker compose up -d minio): write, copy, read back on both paths, plus a missing source, with the threshold and part size lowered to reach the multipart path. There are also transport-level tests, running with no backend, over the copy source, the byte ranges and the storage class. go build ./..., go vet ./... and go test ./... all pass.

[sduchesneau]

Two follow-ups pushed:

The copy source now keeps its slashes literal and escapes each key segment on its own (bucket/root/a%20folder/an%20object). Escaping the whole bucket/key gave %2F separators, which AWS and MinIO decode but other S3-compatible backends do not.

A backend that answers NotImplemented or MethodNotAllowed to the copy — the single call, the multipart start, a part copy or the completion — now falls back to reading and writing the object through the client, aborting the multipart upload first. Every other error, HeadObject failures and access denied included, is still returned as is.

Verified again against MinIO, plus transport-level tests covering the escaping, both fallbacks and the non-fallback error path.