TensorRT-LLM PR #11250: engine build time improved, 'Nemotron Super V3 fp8 model : 290 sec to 260 sec for tp=8'

Merged 2026-02-04. Contains an explicit before/after for the runtime engine-build stage on one model and one parallel setting: Nemotron Super V3 fp8, tensor parallel 8, 290 s before the caching refactor and 260 s after. Vendor self-report of a build duration, so it is an upper bound on what the same build costs now, and it is the only per-model engine-build number this research found published anywhere. Doubles as evidence that NVIDIA treats engine build time as a tracked KPI: sibling issue #11542 (2026-02-17, open) asks to 'Display: 1. Engine build time per model, 2. Engine build time regressions, 3. Slowest engine build' in an internal dashboard.