Keep watching an fd after a poll error on Linux - #791

Fixes #787

On Linux, libuv reports POLLERR on a polled fd as UV_EBADF after it has stopped the watcher. __on_uvpoll_event passed that to UVPoll._fatal_error, which ran the callbacks once, closed the poll and left the closed UVPoll in Loop._polls. After that the fd was no longer watched, a later add_writer() raised RuntimeError: ... the handler is closed, and re-arming from inside the callback did nothing.

This change makes a poll error behave like asyncio's level-triggered epoll:

• _on_poll_error runs the reader and writer callbacks as for a ready fd, then restarts the watcher with the current mask if a reader or writer is still set. The callbacks keep being called until they are removed, as in asyncio.
• If the fd is gone (checked with os.fstat, since restarting a watcher on a closed fd makes libuv abort in epoll_ctl) or uv_poll_start fails, _close_on_error drops the poll from Loop._polls, closes it and reports the error to the exception handler. A later add_reader()/add_writer() then creates a new poll.
• _fatal_error is unchanged and still used for the start/stop failure paths.

Tests: four Linux-only tests in tests/test_base.py, run for both uvloop and asyncio, based on the reproducers in the issue (writer keeps being called after the read end is closed, add_writer after the error with and without remove_writer, and a callback re-arming itself). They fail on master for uvloop and pass for asyncio; with this change all pass. The full test suite passes locally (Linux x86_64, Python 3.12, bundled libuv 1.53.0).

One thing to note: like asyncio, an fd stuck in an error state now keeps calling its callback every iteration until it is removed. If you'd prefer only the smaller fix (dropping the closed poll from _polls and reporting the error), I can trim it to that.