use vmsplice(2) to read/write multiple requests at a time

[tv42]

depends: #33

[tv42]

Oh, wow. After all this time, I come to the realization that there was no such thing as length-prefixed multiple requests in one memory area. I wonder where I even got that idea? Maybe from reading l-k back in the day and seeing a patch that never got accepted. Dang.

vmsplice trickery can still help in decreasing copies, but it's not very interesting for avoiding gc churn. I'm bummed now. Well, dropping maxWrite down for Linux helped there a ton, if that still needs more, next up we'll use "avoid sync.Pool for 2" trick from #120 and I think it was @jacobsa who suggested copying small requests out to avoid pinning a whole large buffer. Also BATCH_FORGET will help with one common request storm.

Right now a pretty simple lots of small requests benchmark says one request roundtrip takes half a millisecond and allocates <4kB (profiling blames context.WithCancel).

BenchmarkCreate-12    20000     647504 ns/op        3591 B/op         44 allocs/op

If you have benchmarks you care about, please try to share them! Feel free to create separate issues for that, to avoid drowning this one in those.

[tv42]

A more flattering, and relevant, benchmark:

BenchmarkLookup-12    200000        181849 ns/op        1096 B/op         14 allocs/op

This avoids the NodeID bookkeeping that Create has, and is just a loop over a stat(2) getting ENOENT (no negative dentry cache).

[tv42]

Yup, vmsplice can't do what was thought of here. io_uring could do something similar, but that depends on so much Go infrastructure that it's not really up to us.