Disable Default Bitmask Compression

[Satrat]

SUMMARY:
This issue came up a few days ago where a user tried to run a sparsified model in vLLM: #45. We compress sparse models into the "sparse-bitmask" compression format by default on save, however this format isn't yet supported in vLLM. I'm updating the save logic to disable automatic sparse compression for now, we can re-enable once this is supported in vLLM

TEST PLAN:
Manual test to confirm sparse models are saved as dense by default:

import torch
from llmcompressor.modifiers.obcq import SparseGPTModifier
from llmcompressor.transformers import SparseAutoModelForCausalLM, oneshot

recipe = SparseGPTModifier(sparsity=0.5)

model_stub = "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T"
model = SparseAutoModelForCausalLM.from_pretrained(model_stub, torch_dtype=torch.float16, device_map="auto")

dataset = "ultrachat-200k"

output_dir = "./test_output_sparse"

splits = {"calibration": "train_gen[:5%]"}
max_seq_length = 512
pad_to_max_length = False
num_calibration_samples = 32

oneshot(
    model=model,
    dataset=dataset,
    recipe=recipe,
    output_dir=output_dir,
    splits=splits,
    max_seq_length=max_seq_length,
    pad_to_max_length=pad_to_max_length,
    num_calibration_samples=num_calibration_samples,
)

Output config.json shows the expected dense format:

  "compression_config": {
    "sparsity_config": {
      "format": "dense",
      "global_sparsity": 0.44059220580610386,
      "registry_requires_subclass": false,
      "sparsity_structure": "0:0"
    }
  },