softgpu: Allow almost flat rectangles to go fast

[unknownbrackets]

Improves Chains of Olympus performance by around 12% on the title screen (from ~64% to ~72%, or so, maybe a bit better.) This allows us to skip the (slow) triangle interpolation and also avoid going over the rectangle twice for each half.

-[Unknown]

[unknownbrackets]

With this on that screen, just for interest, the slowest parts are:

• ~21% T&L (about evenly split between them)
• ~9% clipping + vertex decode, etc.
• ~4% display list processing
• ~1% state field updates
• ~1% debug memory tracking
• Remaining ~64% has some threading overhead in it, but normalizing that (probably overcounting a bit due to CPU emulation, trying to capture the time spent on threads):
  • ~60% triangle (23% interp) vs ~40% rectangle (18% interp), after this change
  • ~19% linear interp for sampling
  • ~24% sampling other than linear interp
  • ~15% drawing pixels

-[Unknown]