nir_to_tgsi: Add support for the fcsel opcodes with included compares.

These map really well to the CMP opcodes on most float hardware.  We don't
do this on native-integers HW, because we don't have an icsel_gt/ge
equivalent in TGSI.

i915g:
total instructions in shared programs: 397005 -> 395430 (-0.40%)
instructions in affected programs: 38814 -> 37239 (-4.06%)
total temps in shared programs: 31824 -> 31768 (-0.18%)
temps in affected programs: 378 -> 322 (-14.81%)
total const in shared programs: 56464 -> 56439 (-0.04%)
const in affected programs: 253 -> 228 (-9.88%)
LOST:   0
GAINED: 36

r300:
total instructions in shared programs: 1193530 -> 1189280 (-0.36%)
instructions in affected programs: 276490 -> 272240 (-1.54%)
total temps in shared programs: 168905 -> 169067 (0.10%)
temps in affected programs: 11486 -> 11648 (1.41%)

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
218 jobs for !14880 with ntt-fcsel-compare in 48 minutes and 35 seconds (queued for 1 minute and 47 seconds)
latest merge request