docs(nvfp4): mixed RtN backward on sm_121 is FEASIBLE — nanochat receipt + web research converge

Two corroborating analyses overturning the over-broad 'datacenter-only' verdict:
- NVFP4-NANOCHAT-RECIPE.md: nanochat already trained on GB10 with
  NVFP4BlockScaling(disable_rht=True, disable_sr=True, override_linear_precision=(False,False,True))
  = fprop+dgrad FP4 (RtN) + wgrad BF16; loss 11.09->6.58 / 22 steps (real receipt).
- NVFP4-GB10-MIXED-APPROACH.md: Brave/Exa/Perplexity/Tavily synthesis — SR-cvt + RHT are
  separable pre-GEMM quant kernels; backward matmul == working forward TN FP4 GEMM. Ranked
  impl plan + numeric gate (dgrad rel-err<=0.20, garbage-grad detector). Implementation in
  progress (port to _nvfp4_route.py).

Author: apstenku123