Initial test enablement for LLM Compressor

[markurtz]

Still needs to be significantly tested, but starting point to play around with for LLM Compressor. This enables the session mixin for control of the training graph, loading of a teacher, and passing through everything within the axolotl training config.

Example config addition on top of the base config here https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/llama-3/fft-8b.yaml:

....

compressor:
  recipe:
    finetuning_stage:
      run_type: train
      pruning_modifiers:
        ConstantPruningModifier:
          targets: '__ALL__'
          start: 0
  recipe_args: {}
  teacher:
    base_model: NousResearch/Meta-Llama-3.1-8B
    load_in_8bit: false
    load_in_4bit: false
    strict: false
    flash_attention: true

after setting up the config file, use the following command:
axolotl train CONFIG.yaml