Support for Stacking Recipes

[Satrat]

First PR towards alternating finetuning/one-shot support. Covers recipe and staging changes necessary to enable running recipes stage by stage between training and one-shot. Previously there was no way to differentiate between stages that had been previously applied and the active stage.

Summary of Changes

• Adding an applied attribute to StageModifiers, used to differentiate which stages are initialized for structure only (already applied stages). If a stage has already been applied, we skip skip it during initialization and updates
• Keep track of applied_stages in the RecipeContainer, so that when we reload the recipe after a stage is added we remember which have already been run
• Adding helper functions to the QuantizationModifier for determining if a module is already quantized. This allows us to stack quantization modifiers as long as they don't affect a common module. If attempting to quantize an already quantized module, we raise an exception
• Update recipe serialization code to add index's to stages with the same name, so they aren't overwritten during recipe export

Testing

Added new unit tests to tests/sparseml/transformers/obcq/test_repeats.py

• Test one-shot SparseGPT recipes are able to be stacked, even if their stage names clash
• Test we error out on clashing quantization modifiers
• Test we allow multiple quantization modifiers as long as they don't overlap

[rahul-tuli]

LGTM!