feat(cosmos3-edge): add Thor AV denoise engine and TeaCache

[LiangSu8899]

Summary

Adds a Jetson AGX Thor implementation of the Cosmos3-Edge AV
inverse-dynamics denoise boundary, together with the official reference
runner, reproducible benchmark harness, correctness gates, and user
documentation.

The no-cache FlashRT path executes all 30 velocity evaluations and delivers
the production optimization result. TeaCache is an explicit optional
approximation layered on top.

Performance

All rows measure the same av_inverse_0 30-step denoise boundary with 15
back-to-back hot iterations. Accuracy is the final action versus the official
eager output. The gates are cosine >= 0.999 and relative L2 < 3%. Both
no-cache engines and the 3- and 2-compute TeaCache schedules were rechecked
with a second seed.

Configuration	Denoise P50	Speedup vs Official	Cosine	Rel-L2	Peak memory (GiB)
Official Cosmos3-Edge (eager)	33.14 s	1.00x	reference	reference	not recorded
FlashRT FP8, no step cache	5.792 s	5.72x	0.999983	0.59%	7.85
FlashRT FP8 + NVFP4 FFN, no step cache	5.456 s	6.07x	0.999707	2.42%	7.54
FlashRT FP8 + TeaCache, 15 computes	2.877 s	11.52x	0.999985	0.56%	7.85
FlashRT FP8 + TeaCache, 6 computes	1.153 s	28.74x	0.999984	0.57%	7.85
FlashRT FP8 + TeaCache, 3 computes	0.576 s	57.56x	0.999986	0.54%	7.85
FlashRT FP8 + TeaCache, 2 computes	0.384 s	86.36x	0.999983	0.59%	7.85
FlashRT FP8 + NVFP4 FFN + TeaCache, 2 computes	0.362 s	91.57x	0.999727	2.34%	7.54

The production claim is the no-cache 5.72x FP8 / 6.07x FP8+NVFP4 result.
Those rows preserve all 30 velocity evaluations and separate the FlashRT
engine gain from step caching.

Pipeline optimization

• Registers a dedicated cosmos3_edge PyTorch/Thor frontend and keeps
  existing model routes unchanged.
• Uses static AV buffers and pre-installs the invariant und-conditioned K/V
  prefix.
• Quantizes generated-stream projections to FP8 E4M3 with calibrated static
  activation scales.
• Adds optional NVFP4 W4A4 FFN up/down projections.
• Fuses residual, RMSNorm, quantization, ReLU2, Q/K normalization, and RoPE
  work.
• Uses FA4 generated-stream attention, a native UniPC scheduler step, and a
  slim 60-row final action layer.
• Captures the complete 30-step denoise loop in one CUDA Graph.
• Includes the official action-only reference runner and fixture-backed replay
  paths for boundary validation.

TeaCache

TeaCache bakes a fixed compute-step set into the 30-step graph. Skipped steps
replay native UniPC while retaining the most recent velocity buffer, with no
new kernel or runtime branch.

These TeaCache numbers are workload-specific benchmark observations, not a
dataset-level task-success claim. Camera setup, robot dynamics, task
distribution, conditioning, and fine-tuned checkpoints can materially change
the valid schedule. Every deployment must qualify TeaCache case by case
against its own checkpoint and task data. The 2-compute rows are aggressive
benchmark operating points; 3 computes is the suggested starting point for
this measured workload.

User and developer impact

• Adds flash_rt.load_model(..., config="cosmos3_edge", framework="torch", hardware="thor").
• Adds a documented official-baseline-first workflow and explicit denoise
  benchmark commands.
• Makes all checkpoint and fixture paths explicit; no host layout is assumed.
• Keeps TeaCache off unless a compute schedule is explicitly selected.
• Keeps the integration isolated to the Thor route; unsupported hardware
  fails during model resolution.

Validation

Build validation used Jetson AGX Thor (sm_110), CUDA 13.0.88, PyTorch
2.9.0a0, and NGC PyTorch 25.09.

• GPU_ARCH=110 build of flash_rt_kernels: passed
• GPU_ARCH=110 build of flash_rt_fp4: passed
• Core and FP4 import plus Cosmos binding inventory: passed
• Cosmos CUDA kernel canaries: 17 passed
• Cosmos routing/reference unit tests: 69 passed, 20 skipped because external
  checkpoint fixtures were not supplied
• Repository install/API/calibration smoke: 86 passed
• Benchmark and baseline CLI help checks: passed
• git diff --check: passed
• Private-path, credential, and local-artifact scan: passed

Scope

The table is denoise P50 in a hot process, not raw-request end-to-end latency.
Official tokenization, video preprocessing, and Wan2.2 VAE encode remain
outside this optimized boundary. VAE encode is the next major end-to-end
target.