Add activation checkpointing verification results and report
Full-model verification (200 iters) on Qwen3-8B W4A16 and Qwen3-30B-A3B
MXFP8 confirms: 85% VRAM reduction on MoE with zero accuracy loss.
Includes experiment scripts, lm-eval results, and concise report.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: yiliu30 <yi4.liu@intel.com>
