radv: Properly link TCS->TES IO again.

Instead of relying on driver locations, let's use a prefix sum of the IO mask which also includes the TES inputs read masks.

This fixes the small regressions from !28488 (merged) and makes RADV TCS->TES IO great again.

Fossil DB stats on Rembrandt are all-green:

Totals from 2634 (3.32% of 79395) affected shaders:
MaxWaves: 56336 -> 56450 (+0.20%)
Instrs: 1666290 -> 1663744 (-0.15%); split: -0.22%, +0.06%
CodeSize: 8653432 -> 8621152 (-0.37%); split: -0.39%, +0.02%
VGPRs: 126776 -> 126608 (-0.13%)
IN: 30910 -> 25426 (-17.74%)
OUT: 27000 -> 21516 (-20.31%)
LDS: 10382848 -> 10556416 (+1.67%)
Latency: 9393986 -> 9382070 (-0.13%); split: -0.25%, +0.12%
InvThroughput: 2125714 -> 2125417 (-0.01%); split: -0.13%, +0.12%
VClause: 33477 -> 34396 (+2.75%); split: -0.31%, +3.05%
SClause: 27883 -> 27865 (-0.06%); split: -0.27%, +0.20%
Copies: 82262 -> 82702 (+0.53%); split: -0.34%, +0.87%
PreSGPRs: 91002 -> 90987 (-0.02%)
PreVGPRs: 108505 -> 108473 (-0.03%); split: -0.06%, +0.03%
VALU: 1086352 -> 1085970 (-0.04%); split: -0.18%, +0.14%
SALU: 181781 -> 178469 (-1.82%); split: -1.83%, +0.01%