Bringing laplace-torch to foundation-model era

[wiseodd]

Main features of this pull request:

1. Support only doing Laplace on params that require grad. Use case: PEFT (like LoRA) on top of a frozen foundation model. This is more efficient than SubnetLaplace since the latter still computes the full Jacobians.
2. Add support to multiple leading dims in classification likelihood. E.g. the logits is (batch_size, seq_len, n_classes). Useful for language modeling and reward modeling.
   1. This PR also contains the integrate-latest-asdl changes. I tested it with my ASDL fork (only a couple of light changes to support weight-sharing dim and ignore_index; so crucial in language modeling): https://github.com/wiseodd/asdl/commits/dev/. Please also check this and let me know what's the most elegant way.
3. Support Huggingface dataset. The assumption is that x is a UserDict containing input_ids, attention_mask, etc., things that are produced by HF dataloader.
4. Add a new likelihood called reward_modeling where the classification likelihood is used during training and the regression likelihood is used during prediction.
5. Add support to torchmetrics for gridsearch. The benefit is that it supports running metrics => less memory overhead (vis-a-vis gathering all the logits first).
6. Add Jacobian computation with torch.func (functorch, really) as a general Jacobian computation for GLM predictive. Useful for Bayesian optimization/invariance learning where you need to backprop through the variance. Much more elegant than to change ASDL.

Relevant unit tests are provided. All tests passed; the only ones failed are the old LowRankLaplace issues.

[wiseodd]

I replaced get_nll in crossval with RunningNLLMetrics(). So this PR will close #160 .
(I thought before that I RunningNLLMetics hadn't been implemented, so would have involve major work.)

[wiseodd]

Might as well fixes #156 while we're at it.

[wiseodd]

All tasks finished!

[wiseodd]

Double-checked and everything looks fine! Merging this.