fine-grained ME distribution

[Steve Borho]

In medium preset and below, when a CU is analyzed for inter there are up-to fifteen motion searches that need to be performed and merge/bidir candidates to measure:

2Nx2N L0-0, L0-1, L0-2, L1-0, L1-1, merge, bidir Nx2N " 2NxN "

And there are potentially asymmetrical partitions as well. These could each be done by an idle worker thread. Infrastructure should be added to allow this work to be distributed. This might remove the need for WPP to keep cores saturated in many cases, recovering that 1% compression efficiency cost.

[Steve Borho]

this is currently in the public repo; along with fine-grained mode decision (--pme and --pmode respectively) but both will be seeing more improvements over the coming weeks. --pme in its current sense is rarely a performance gain and --pmode helps most at --rd 4