Add enterprise GPU benchmarks and clarify FP8 model card
Populate single-GPU Generator throughput for DGX Spark, DGX Station, and RTX PRO 5000 72GB Blackwell from enterprise PBR 309485. The six FPS values use 189 frames divided by measured average generation time and carry an explicit assumption footnote; the underlying seconds and unverified denoising-step count are documented alongside the table.
Clarify Reasoner video inputs and throughput scope, correct H100/H200 hardware descriptions and action array shapes, document resolution-specific video frame limits and the configured mixed activation precision, and distinguish base-model quality plots from FP8 evaluation. Fix asset downloading, framework links, stale SGLang modality wording, and Explainability output/benchmark labels.
Validation: 14 existing Generator FPS values and 18 Reasoner values match the supplied source workbooks; six enterprise generation times match PRISM report 40954 and its workbook. Checked all six derived enterprise FPS cells, Python example syntax, example JSON, and git diff whitespace. GPU inference and backend loading were not run. Existing DROID policy values still require checkpoint/precision provenance; the GB300 policy estimate remains labeled.