TensorRT-LLM PR #7633: '[None][infra] Disable CU12 build to save build time (cost > 5 hours on SBSA)'

Merged/closed PR from 2025-09-09. The single most direct statement of TensorRT-LLM's build cost found anywhere: the CUDA 12 build leg of the project's own CI cost MORE THAN 5 HOURS on SBSA (Grace Blackwell, aarch64 server), and the fix was to delete that build leg rather than make it faster. Internal-infrastructure PR - it leaks NVIDIA's own CI cost, which is why it is worth citing even though the number is a vendor-internal self-report of a cost rather than a performance claim.