Bound the two-pass wet-pass mode search with the dry pass's rd

[yeqing-wu]

The wet pass starts mode search from an unbounded rd limit, so its pruning cannot bite until it finds a good candidate on its own. Start it instead from the dry pass's rd for the same block and size, plus a margin.

The two passes see different contexts, so the bound can be violated. A block left with no valid mode is re-searched once with the original limit.

Anchor: commit c77f169
Speed 4 (cpu-used=4): FG16 CTC (33 frames, class A1 and A2, RA)

  +------------+-------+-------+-------+-------+------+------+
  | Class      |     Y |    Cb |    Cr |  wAvg | Enc% | Dec% |
  +------------+-------+-------+-------+-------+------+------+
  | A1         | -0.30 | -0.29 | -0.38 | -0.30 |  100 |  100 |
  | A2         | -0.17 |  0.07 |  0.31 | -0.15 |  101 |  100 |
  | Avg w/o B2 | -0.21 | -0.04 |  0.11 | -0.19 |  100 |  100 |
  +------------+-------+-------+-------+-------+------+------+

STATS_CHANGED

[yeqing-wu]

@urvangjoshi, Update on reusing the MV from the dry pass as the starting MV of the wet pass: the gain has dropped from -0.13% to -0.03%. Part of the reason is that the anchor itself increased an additional -2% gain compared to before.

Meanwhile, there have also been a lot of recent commits (several for MV search and DRL), so I'm still going through them to check whether anything was missed, and to see if I can push the gain further. Just let you know.