perf: Reduce memory overhead for multipart uploads

[kylebarron]

As described well in #767, we currently always copy input into a scratch buffer for any pull source. But for buffer sources this is unnecessary as we can upload the input buffer directly.

Change list

• Make Buffer a first-class variant of PutInput, rather than shoe-horning it into a PullSource. Also we no longer need to use a Cursor wrapping the Buffer.
• Don't use a scratch buffer copy when uploading a buffer
• For a pull source, create a single scratch buffer outside the loop, which we reuse on each iteration, rather than creating a new buffer allocation in every loop iteration.

This PR was created in discussion with Claude but I wrote the source code changes myself. Claude wrote the new tests.

Closes #767