- yield was in wrong if block, not sure how

- count ULTs in progress pool that are waiting on communication as
  waiters so that we don't busy spin when it isn't necessary
- if a trigger occurs, yield immediately to service ULTs that may have
  been activated as a result
- HG_Progress() is still called too frequently, see TODO note
- outcome: this makes servers significantly faster, but seems to make
  clients a bit slower; optimal performance right now comes from
  building servers with this git commit and building clients without
  it; need to investigate

in cases where margo_forward_timed calls timeout, margo will
return HG_TIMEOUT rather than HG_CANCELED

- this doesn't cause a warning and makes the prototype a little less
  confusing now in conjunction with previous commit

- the mercury hg_handle_t is really a pointer that references an
  internally defined structure, we can (and should, for proper
  reference counting purposes) pass it directly to ULTs without copying
- code using margo must be updated accordingly, handler ULT argument is
  an hg_handle_t now, not an hg_handle_t*

- first test is for the sleep functionality

we need to remember whether a sleeping thread is in the same pool
as the progress thread to properly decrement the number of
waiting ULTS.