Add activation checkpointing benchmark scripts and findings
Includes investigation scripts, micro-benchmarks, full-model
verification log, and FINDINGS.md documenting:
- Why block-level checkpointing fails for MoE (peak moves to backward)
- Per-WrapperLinear checkpointing solution (85% VRAM reduction)
- Full Qwen3-30B-A3B verification: 11.74 GB peak, correct generation
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: yiliu30 <yi4.liu@intel.com>