Bring in a bunch of well tested MPIPE changes. Preallocate a minimum

Bring in a bunch of well tested MPIPE changes.  Preallocate a minimum
number of mpipe elements when it is initialized.  Use an array to cache
free MPIPE buffers nad remove the data structure overloading that was
previously occuring on the buffer itself.  Add a deconstructor.  Separate
the blocking and non-blocking allocation APIs into their own functions.

The new code still needs Giant, but it's getting a lot closer to being
lock free.