Data Batching, precompilation, and autotune

[kazewong]

In this PR we implemented a couple of functionailties:

1. Precompilation: Now the kernels and updates are constructed and compiled outside the make_xxx functions, inside a function precompilation in LocalSampler_Base.py. Previously there are multiple copies of likelihoods, kernels and update in the code, which bloated the compilation time. Now there should only be two compilation. This should address the concern in Precompile kernel #111
2. Moved autotune for MALA inside the MALA class. In previous version of the code, MALA autotune would makes its own vmap kernel, which would trigger recompilation. This is now handled properly so no recompilation should be needed.
3. Data batching. Now the sampler can take some additional data. This should address the concern in Parsing arguments other than x into log-posterior probability functions #93.

There are a couple more to-do items before we can merge this PR:

• Modify random walk Gaussian
• Modify HMC
• Update examples
• Update documentations
• Update tests

[kazewong]

Basic documentation on the top level are restructured. Specific tutorials and more example needs to be refactor too. Merging the changes to main and release the latest version for now.