Make socket "accept" methods dont block fibers (Fiber::Scheduler support) - #9813

The accept (and sysaccept for servers) methods of ServerSocket, TCPServer and UNIXServer were blocking the fiber threads until some client connects, never calling the io_wait.

When a client in another fiber tries to connect on, using the same scheduler, it never run, blocking the thread permanently.

Now they try a nonblocking accept, and while nothing is pending, wait on scheduler io_wait without holding any lock.
The io_wait result is handled as in "rb_io_maybe_wait_readable", and the server closing interrupts a waiting fiber as in #9760. Without a scheduler, the extra cost is one FiberScheduler.current call.

It also fixes 3 existing bugs around these methods:

• ServerSocket#accept raised IO::EAGAINWaitReadable once IO.select had left the channel nonblocking
• UNIXServer#accept was returning a not connected socket when another thread or fiber took the connection first
• The GC was closing the descriptor returned by UNIXServer#sysaccept when it finalized the socket it came from

The new specs passes on MRI 3.4 and 4.0. This test test/async/grpc/client.rb of async-grpc was hanging.
But now it passes together with #9805 of @sampokuokkanen.