Throwing cores at the problem!

[unxmaal]

Changes:

• Async compiler thread to improve throughput. ~25% boost from unblocking interpreter
• Batched cp0_count via step_lite() for JIT interpreter bursts == ~40% boost from removing
  per-instruction overhead
• Flush-reload pattern in load helpers. Removes the 3-helper ceiling.
• NOTE: the regalloc2 bug turned out to be a phantom
• Verify mode improvements. Reduces false positives from 122K to 547 (all MMIO)
  • branch-likely step count
  • MFC0 timing
  • exception detection during re-run
• Dormant write log infrastructure for future store speculation work