There are several points where ByteStreamUploader may discover that the server already has the blob fully uploaded. These points tried to effect an early return from the upload code, generally by "lying" to higher layers that the upload fully finished. That could lead to bugs. For example, consider the added test case: A compressed upload completes its writing but receives an error instead of a server ACK. On retry, QueryWriteStatus reveals the blob exists and returns its uncompressed size. This confused the checkCommittedSize logic, which expected the final committed size of a compressed upload to be the total compressed data size or -1. The code added by bazelbuild@daa3dbe also looks broken in the case of compressed uploads.

Rework the uploader code, so that early returns throw a AlreadyExists exception. The exception control flow naturally reflects the desire to escape quickly to the top level.

Closes bazelbuild#17791.

PiperOrigin-RevId: 517389227
Change-Id: I23a2ae92fd4ad27dad750418c128c0d0b245e573