Make os.zig not depend on the event loop

[kristoff-it]

This is #5620, I force-pushed and it broke the PR.

Currently os uses the event loop when receiving EAGAIN, which doesn't allow us to support completion APIs like io_uring. With this PR, now it's fs and net that (in async mode) delegate the whole I/O operation to loop.zig, and it's loop that can then decide to do the readyness dance with EAGAIN, or (once someone implements it) use io_uring or equivalents.

Read the comments in #5620 for more details about sendfile and open design questions.

[kubkon]

Re Windows CI erroring out, I think the error code of 3221226505 that we might see, and/or the error message LLVM ERROR: out of memory are correlated with Azure live site incident currently ongoing. This is me just guessing right now since I can't replicate any of those on my local Win install, not to mention the fact that #6397 which had changes pushed today passed Win CI just fine.

[FireFox317]

One thing to consider is that the semantics of the functions in std.os are not valid anymore when using async io, because then an error (WouldBlock) will be returned, while in reality there really is no error.

However it is significantly better to not have any event loop code in the std.os functions.

[kristoff-it]

@FireFox317 yes, the idea is that if you're using high-level APIs (i.e. fs and net), you can continue doing so and nothing should change, while if you want to use lower-level APIs, then you have to either call into the event loop or call into os but then handle WoulBlock manually.

[andrewrk]

Don't forget the doc comments in os.zig

[kristoff-it]

Docs updated and it didn't break the build (you never know :D)

[andrewrk]

This is ready to land. I have 1 little fixup for you that doesn't actually change the behavior.

[codehz]

maybe typo? I/0 vs I/O