fix minor issues with lm_decomposition.py pipeline

[mivanit]

Description

1. bump torch dependency to require >2.4.1
2. add weights_only: bool kwarg to lm_decomposition.py:main()

Motivation and Context

1. because pytorch <=2.4.1 does not have nn.Module.set_submodule(), running lm_decomposition.py gives the error:

  File "~/apd/spd/experiments/lm/lm_decomposition.py", line 146, in main
    optimize(
  File "/home/miv/projects/MATS/apd/spd/run_spd.py", line 242, in optimize
    layerwise_random_recon_loss = calc_layerwise_recon_loss(
                                  ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "~/apd/spd/losses.py", line 135, in calc_layerwise_recon_loss
    modified_out = model.forward_with_component(
                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "~/apd/spd/models/component_model.py", line 141, in forward_with_component
    self.model.set_submodule(module_name, component)
    ^^^^^^^^^^^^^^^^^^^^^^^^
  File "~/apd/.venv/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1729, in __getattr__
    raise AttributeError(f"'{type(self).__name__}' object has no attribute '{name}'")
AttributeError: 'GPTNeoForCausalLM' object has no attribute 'set_submodule'

2. due to pytorch vuln https://nvd.nist.gov/vuln/detail/CVE-2025-32434 huggingface transformers library doesn't allow loading with weights_only=True, adding a kwarg for this in lm_decomposition.py:main() which is passed on to load_pretrained()

Does this PR introduce a breaking change?

Probably not. Most likely to cause issues is the bumping of the pytorch dep.