fix: broken get and replace for s3 client

[sandeepsukhani]

What this PR does / why we need it:
When GetAndReplace callbacks return opaque io.ReadCloser wrappers (such as io.NopCloser), objstore.TryToGetSize cannot determine the content size, causing minio-go to use a multipart upload path that silently reconstructs PutObjectOptions before CompleteMultipartUpload, dropping customHeaders and, therefore, the If-Match conditional write header. S3 then accepts the upload unconditionally, allowing a stale-ETag write to overwrite a newer version of the object.

To fix it, I have added sizedGetAndReplaceBucket, which wraps the caller's callback inside GetAndReplace and buffers the returned reader when its size is unknown, replacing it with an objstore.NopCloserWithSize-wrapped bytes.Reader. This gives minio-go a known content length, forcing the single-part PUT path where If-Match is preserved end-to-end.

GetAndReplace is only used to update TOC files, which are not that large, so switching them to a single-part upload should not cause any issues. We can try to fix the issue upstream or change the reader interface to match what objstore.TryToGetSize handles.