vc4: Jump to program end when all fragments are discarded

[anholt]

If a whole 16-pixel area is discarded, we could jump to the end of the program and skip the rest (particularly any remaining texture lookups). Intel saw pretty big wins from their corresponding change to do this.

We may or may not see any improvement from also conditioning texture fetches on any of the pixels in each 2x2 subspan having been not discarded.

Potential apps:

• glbenchmark 2.7: Jumps over texture sampling
• mupen64plus: Big jumps over control flow with math. Interesting case of conditional discard within control flow near the end of a program.
• glmark2: No benefit but make sure it doesn't regress (short jumps).
• minecraft: No benefit probably, but make sure it doesn't regress (medium length jumps).

[anholt]

I've started on this one

[anholt]

Branch is up at vc4-discard-jump. Didn't get any measurable wins from it on 3DMMES.