optimize disk copy for multipart uploads

[sbrudenell]

with #51, we copy from the send stream to disk for multipart uploads. we currently do this in python code, which is slow. we could implement a faster copy using sendfile() and splice().

we could also reuse buffers with readinto() for the slow case.

Some notes:

• sendfile() requires mmap-capable input, so won't work for our normal case
• splice() requires 3.10+ and a capable filesystem