Major bugfix + new efficiency-focused functionality.

[patrick-kidger]

Heyhey. I've been tweaking some of the SDE-GAN stuff (for the paper, not the example), and realised that it could be made much more efficient with library support... and what's the point of being a developer of the library if not for that?

• Major bugfix: fixed calling solver.integrate twice in sdeint.py::integrate (!) (I can make this a separate PR if you want to put this in immediately.)
• Added f_and_g, g_prod, and f_and_g_prod as functions that the SDE can define, which will be used in the solvers if possible. f and g no longer actually have to be defined if the solver doesn't need them. [1]
• Improved efficiency:
  • logqp=True now takes advantage of this to define an f_and_g function that is more efficient than calling them separately. (Unrelated to what I'm doing, but why not.) [2]
  • Milstein and derivative-free Milstein now no longer have extra g computations.
• Improved the error checking in check_contract to explicitly require a single batch dimension.

(N.B. This PR depends upon #85.)

[1] My use case for this: First of all, I'd like drift and diffusion to just be different final layers at the end of the same MLP, hence f_and_g. Second, for the discriminator, the diffusion ends up being a product of matrices. Doing (matrix-matrix)-vector would be inefficient, hence g_prod/f_and_g_prod.

[2] There's still some improvements that could be made here: for simplicity logqp still requires f/g/h to all be specified, without exploiting f_and_g or f_and_g_prod if they're available. The logical extension of this would be to also add f_and_g_and_h, f_and_g_prod_and_h as possible methods.

[lxuechen]

Thanks for this update! solver.integrate twice is a serious bug that might have been introduced by merging commits. Could you split this into a separate PR? We could merge this to master immediately.

Could you also split bullets 2-4 into 2 separate PRs? One for 2-3, and another for 4. That'll help a lot in terms of going through the code.

It might be too pesky to split this into 4 PRs, but I think it's reasonable to factor out the f_and_g functionality, as I think it would affect a few things.

[lxuechen]

For efficiency purposes, I'm committing minor tweaks so we can merge. For the record, I'm rebasing from dev directly after merging #85, and I added the license for diagnostics/run_all.py