Fix recently introduced codepaths that are involved in parallel
submodule operations, which gave up on reading too early, and
could have wasted CPU while attempting to write under a corner
case condition.

* sb/submodule-parallel-fetch:
  hoist out handle_nonblock function for xread and xwrite
  xwrite: poll on non-blocking FDs
  xread: retry after poll on EAGAIN/EWOULDBLOCK