Align OG2 FP8 recipe with PR #1500 layer precision approach

Replace broken fp8_first_last_bf16 mechanism with resolve_layer_precision()
from PR #1500. The old approach set config attributes that were never read by
the forward pass, causing all layers to default to FP8 regardless of setting.

Key changes:
- Delete fp8_debugging.py, add quantization.py with resolve_layer_precision()
  and initialize_quant_stats_logging()
- Add set_recipes()/get_layer_autocast() to OG2 model (from lepton branch),
  model now handles per-layer autocast internally
- Model constructor accepts fp8_recipe/fp4_recipe, set_recipes() called after
  FSDP wrapping since recipes aren't serializable
- Remove outer te.autocast() from training loop (model handles it)
- Rename fp8_stats_config -> quant_stats_config throughout
- Add _parse_layers_cfg() for CLI string support
- Add og2_7b_fp8_fl1_pq2.yaml with explicit fp8_layers=[2..31]
- Expand fp8_debugging_stats.yaml with all layer types + LogTensorStats

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: savitha-eng