Feature: TransformerEngine FP8 support

[sfc-gh-mwyatt]

Adding support for FP8 autocasting feature from Transformer Engine library.

• Replace nn.Linear layers with transformer engine linear layer in loaded models for layers that substring match fp8_target_modules
• wrap loss calculation with autocast
• Add config support for fp8_recipe:
  • requires specifying a type that maps to Recipe subclass
  • Requires specifying fp8_target_modules

Example Usage:

model:
  name_or_path: Qwen/Qwen3-8B
  fp8_recipe:
    type: delayedscaling
    fp8_format: hybrid
    amax_history_len: 16
    amax_compute_algo: max
  fp8_target_modules:
    - q_proj
    - k_proj
    - v_proj
    - up_proj
    - down_proj
    - gate_proj

data:
  sources:
    - type: huggingface_instruct
      name_or_path: HuggingFaceH4/ultrachat_200k:train[:1000]

[sfc-gh-mwyatt]

Perhaps this should be named differently. TE supports both FP8 and FP4 - plus there are other libraries we may wish to support. A better name may be:

• quant_recipe
• quant_target_modules

[sfc-gh-mwyatt]

@sfc-gh-jrasley @sfc-gh-zhyao thoughts?