Generalize autodiff functions

[lyndond]

Info:
Methods vector_jacobian_product and jacobian_vector_product can now take multiple vectors (useful for so-called block iterative methods).

Summary:

• vjp and jvp can take blocks of vectors and compute products along each direction
• fixed bug in jacobian(); handles edge-case where dim_y = 1 and it returned a flat (dim_x,) instead of (1, dim_x) tensor
• jacobian() issues a warning when input dimensionality is >1E6
• tidied up documentation for each function
• removed some user-facing control to handle creating/retaining graph during vjp/jvp
• vjp now has optional Bool argument to detach output
• created tests for all autodiff functions in test_synthesis.py
• functions in eigendistortion.py were changed accordingly to accommodate new changes

[billbrod]

I'll look at this this afternoon, but this is what's in #10 , right?

[billbrod]

Some small changes requested, largely on docs

[billbrod]

Yeah and just to reiterate my question from earlier, does this accomplish what I put in issue #10 ? That is, can they now accept 3d/4d inputs or is the reshaping trick still necessary?

EDIT: Answering my own question, no they do not. That's fine for this commit, but should think about whether we want that eventually.

[billbrod]

I added some small comments, overall this looks good to me. Tutorial in particular looks good.