Implement zero-copy writes for the multiplexer

[bdraco]

closes #394

Note that this likely won't make any difference until #369 and #344 because the current implementation forces many small reads which negate much of the benefit here. Since we benchmark both sides of the connection, the benchmark won't improve much either until we can get the read buffer size a bit more aligned with the use case.

On 1 MiB message this will make quite a bit of a difference for a write op even if its only a small part of the whole stack:

[codspeed-hq]

CodSpeed Performance Report

Merging #395 will not alter performance

_{Comparing bdraco:zerocopy_writes (85f254a) with main (506c0eb)}

Summary

✅ 4 untouched benchmarks

[bdraco]

Marking as draft to wait for #369 and #344 since it will change the dynamic