Improved poll and efficient epoll

[nyh]

Right now poll() has a rather unusual design favoring the performance of immediately-returning poll (we poll_scan() before trying to install the poll request on the descriptors). We also do not efficiently support epoll, and epoll_wait() just calls poll() which is terrible for performance.

Instead of the current poll.c code, let's design our own osv::poll object, inspired by epoll. It will have an add() method to install itself on one file structure, remove() to uninstall, and wait() to wait on any of these files (or return immediately). We will wait using a "waker" (our familar wait_until+threadid idiom) - I believe we don't need condvar (we only ever have one waiter in a poll request) and certainly not msleep, and the "waker" idiom also takes care of wake-before-wait races without needing another variable or a mutex to protect it.

osv::poll will also have a static function osv::poll::wake(file *fp) to wake any polls sleeping on this fp (and also set their revents and add them to a list of ready files). The "extern C" function poll_wake() will just call osv::poll::wake.

The epoll code will wrap osv::poll. The poll() code will temporarily create an osv::poll (just like we have a poll_request in the current code), add() to it the given fds, call wait(), and finally prepare the poll results.

[nyh]

I'm closing this issue - we have a completely rewritten epoll now, and it's no longer a wrapper around poll, solving the original problem (although with a different design than what I suggested above).