Reduce object body memory copies

[louisgls]

Summary

• avoid cloning successful S3 response bodies in the raw-response middleware
• preallocate request and response body buffers when Content-Length is known
• drop plaintext/ciphertext references earlier after encryption/decryption to reduce peak live memory
• add unit coverage for fixed-size body reads and short-body errors

Context

This is a focused upstream version of the memory work discussed in #28 and in the k3s-argocd-cluster/s3proxy fork. The fork contains additional unrelated features such as request caching, B2/no-tagging support, chart edits, and workflow changes; this PR intentionally keeps only the object-body memory reduction.

Unlike #28, the raw-response middleware still stores the error body in metadata after reading it, so forwarded S3 error responses keep their raw XML body.

Validation

• go test ./...
• CGO_ENABLED=0 GOOS=linux go build -o /tmp/s3proxy-memory ./s3proxy/cmd