Add MACE-MH-1 + MLP benchmark for matbench_mp_e_form

[mark-alence]

Summary

• Frozen MACE-MH-1 invariant descriptors (1024d) + calculator features (19d) + Magpie composition features (132d)
• 3-layer MLP head (256→128→1, SiLU, LayerNorm, dropout=0.1)
• matbench_mp_e_form MAE: 0.0169 eV/atom

Method

1. Extract mean-pooled invariant descriptors from pretrained MACE-MH-1 (mace_mp(model="mh-1", head="omat_pbe"))
2. Extract energy, force, and atom energy statistics from MACE forward pass
3. Extract Magpie composition features via matminer
4. Concatenate all features (1175d total), train MLP with AdamW + CosineAnnealingLR + L1 loss for 200 epochs