fix: reduce threads per block for FRI tracegen launcher

[stephenh-axiom-xyz]

Resolves INT-4950.

Some CUDA 13.0 optimizer choices caused kernel fri_reduced_opening_tracegen to use 80 registers (previously 58), causing the total number of registers per block to exceed the allowable amount set by the device.

Benchmark difference on CUDA 12.9 is negligible. Benchmarks done to determine savings gained by switching to CUDA 13.0 will be done later.

[github-actions]

group	app.proof_time_ms	app.cycles	app.cells_used	leaf.proof_time_ms	leaf.cycles	leaf.cells_used
verify_fibair	(+31 [+11.6%]) 298	322,610	2,058,654	-	-	-
fibonacci	(+12 [+1.2%]) 1,029	1,500,210	1,060,232	-	-	-
regex	(+37 [+1.3%]) 2,967	4,108,483	13,403,604	-	-	-
ecrecover	(-4 [-0.4%]) 958	140,497	2,275,066	-	-	-
pairing	1,546	1,882,939	23,722,622	-	-	-

Commit: 0002889

Benchmark Workflow