The previous implementation used only one BUDDY_REQS, so whenever one
segment allocation was fulfilled, other requests with lower priority
could "get through" and ultimately lead to bfa_alloc() failing to
complete.

By using two BUDDY_REQS, we now make sure to "keep out place in the
priority queue".

We also limit cramming not only by the available bitfield segment slots,
but also by a maximum of 4 (1/16th of the requested size) and yield when
a lower cram does not succeed to motivate LRU more to make room.

This has undergone a _lot_ of testing and has gone through many
iterations, which all have been squashed into this commit.