Cleanups for the socket locking mechanism.

Improve RPC request ordering by ensuring that RPC tasks that are
already queued on xprt->sending get sent before tasks that happen to
get scheduled just when there is a free slot.

In case the socket send buffer is full, queue the tasks on
xprt->pending rather than xprt->sending in order to eliminate the risk
of accidental wakeups from xprt_release_write() and xprt_write_space().