feat: add torch_compile option

[othertea]

This PR adds an option to use torch.compile through the appropriate argument in HuggingFace's TrainingArguments.

In my experiments, with recent versions of PyTorch and Transformers, this resulted in faster training and slightly lower GPU memory usage; more details about my experiments can be found in the discussion in #173. With this change, we can simply add torch_compile: true in the config files in chemnlp/experiments/configs under the trainer section to use torch.compile, for instance, in the following line:

chemnlp/experiments/configs/hugging-face/160M_full.yml

Lines 30 to 31 in 97d6b3f

	per_device_eval_batch_size: 4

[jackapbutler]

thanks!