Add Gemma 4 E2B MLP quantization and multi-pass sweep configuration.

[copybara-service]

Add Gemma 4 E2B MLP quantization and multi-pass sweep configuration.

• Support YAML sweep specifications in xm_launch.py with macro expansion, dry-run support, and per-trial output directories.
• Support hierarchical scaling and multi-pass quantization configurations for MLP layers in MaxText (base.yml, types.py, quantizations.py).
• Add gemma4_e2b_mlp_quant_sweep.yaml defining 13 parallel trials on Gemma 4 E2B (Ghostfish 2x2x1):
  • Trial 1: Baseline (unquantized BF16)
  • Trial 2: 1-pass mxfp8_16 on all MLP tensors (flat scales)
  • Trial 3: 1-pass mxfp8_16 on all MLP tensors + hierarchical scaling on weight gradients (drhs_mlp_grad)
  • Trial 4: 1-pass mxint8 on all MLP tensors (flat scales)
  • Trial 5: 1-pass mxint8 on all MLP tensors + hierarchical scaling on weight gradients
  • Trial 6: 3-pass triangular mxfp8_16 (all MLP, flat scales)
  • Trial 7: 3-pass triangular mxfp8_16 + hierarchical scaling on weight gradients
  • Trial 8: 4-pass full-cross mxfp8_16 (all MLP, flat scales)
  • Trial 9: 4-pass full-cross mxfp8_16 + hierarchical scaling on weight gradients
  • Trial 10: 3-pass triangular mxint8 (all MLP, flat scales)
  • Trial 11: 3-pass triangular mxint8 + hierarchical scaling on weight gradients
  • Trial 12: 4-pass full-cross mxint8 (all MLP, flat scales)
  • Trial 13: 4-pass full-cross mxint8 + hierarchical scaling on weight gradients
• Add launch_gemma4_e2b_mlp_quant_sweep.sh using pretraining reference settings from cl/980100775 (Grain dataset loading, Ghostfish 2x2x1, 10k steps, BS=4, GA=16, 1.05M tokens/step, /cns/yo-d/ storage).