Better integration with libuv's event loop when using MPI as transport

[amitmurthy]

Currently we do a busy-wait with MPI.IProbe - yield() loop, which consumes CPU cycles unnecessarily.

[eschnett]

This is actually difficult to solve, since the MPI implementation probably needs to poll the networking hardware periodically. Integrating the MPI implementation with libuv may make sense, but I don't know of an MPI implementation that does this, or that offers a select-like API function.