perf(gpu): tail-path kernel for jagged constraint poly eval zerocheck

[erabinov]

Summary

• New zerocheck_eval_tail.cu CUDA kernel that exploits parallelism over evaluation points in the later (tail) rounds of zerocheck, where the per-point workload shrinks and standard per-variable parallelism leaves the GPU underutilized.
• Threshold-based dispatch in sp1-gpu-zerocheck to switch into the tail kernel once the remaining variable count crosses the threshold.
• Wiring in sp1-gpu-sys (v2_kernels.rs, CMakeLists.txt) to build and expose the new kernel.
• BlockAir implementation for GlobalChip (14 blocks: Poseidon2 permutation rounds mirroring Poseidon2WideChip, separate blocks for the curve-formula check, y6 sign check, sum_checker_x/sum_checker_y, and a dedicated block for all interactions). RiscvAir::Global is now dispatched through the per-block path. This increases the parallelism of constraint-polynomial evaluation on the GlobalChip and gives a further speedup on top of the tail-kernel work.