revert: remove PMAT-511 per-layer backward (dead end, wrong-direction gradient)

Per-layer backward through W_down^T @ W_gate^T without SiLU derivative
injected WRONG-DIRECTION gradient that made training WORSE:
- Without mixing: epoch 1 loss 2.97, final 16.11
- With alpha=0.1: epoch 1 loss 3.63, final 17.77 (WORSE)

Reverted to simple same-gradient approach. Path B (cuBLAS hybrid via
--gpu-backend cuda) replaces this — the CUDA backward in backward.rs
has proper per-layer backward with RMSNorm, SiLU, and attention.

Refs PMAT-511

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: noahgift