sysfs: implements PollRead with poll instead of select

[evacchi]

In all usages, it looks like poll should be preferred to select: the API is simpler and easier to emulate on other platforms; above all, it does not use FdSets and won't change pointers in-place.

We remove all usages of select and FdSet and instead expose and use poll and []PollFd, which is comparatively much more straightforward. We also rewrite the select emulation on Windows, and instead emulate poll, based on the same principle: we use WSAPoll for sockets, PeekNamedPipes for pipes, and report all regular files as "ready".

This in turn should eventually make it possible to remove File.PollRead or at the very least simplify further poll_oneoff.

[codefromthecrypt]

I messed up the other API, signature wise (duration not int64) and will fix in another PR.

Meanwhile, this looks good! I was verbose about style stuff so that we can try to be coherent with other style by default. This helps the project grow and feel to users as if it was written by one person, even if that's never the case.

Cheers.

[codefromthecrypt]

I opened #1597 to rebase the existing implementation, which will drift this. If it merges today, I will rebase your PR!

[codefromthecrypt]

OK I reworked this to pare down the change around implementing PollRead with poll instead of select, which is the main part. Main thing I think waiting here is to revisit the RATIONALE which is about select instead of poll. More work can happen in future PRs, but be wary of platforms that don't support syscalls such as plan9. More on that below!

A future change can implement a multi-poll, but that wouldn't seem to make sense on a file because it has only one file descriptor. In fact I would probably suggest not doing this until forced to, because there are a lot of edge cases we have to test. Even here we have logic to handle many types of inputs, but the only one tested is a single fd on read. In other words, unsolicited 2p is to not let the impl grow too far in front of reqiuirements and lazy grow the impl only when tests and ABI moves to require them.

[codefromthecrypt]

build break seems unrelated

 /usr/bin/sh: line 1: osslsigncode: command not found
mingw32-make: *** [makefile:302: build/wazero_windows_amd64/wazero.exe.signed] Error 127

[evacchi]

further work would be generalize the use of poll in poll_oneoff (right now it is only used for stdin). I am now updating RATIONALE.md

[codefromthecrypt]

RATIONALE LGTM