Batching L-BFGS-B (using as much as possible from scipy's implementation) (#2870)

Summary:
Pull Request resolved: #2870

This diff takes the python frontend of L-BFGS-B copied to our codebase previously and extends it to support batched functions.

The following points are important:

- It does the batching by calling the C/FORTRAN code running l-bfgs-b in scipy in a for loop on each iteration
- It makes the assumption that the to be optimized function is truly batched in the sense that each item in the batch is processed completely independently and they do not have any effect on each other. This is used as the batch will shrink over time, as more items finish optimization.

The ideal long-term outcome would be to upstream these changes to scipy.

Reviewed By: Balandat

Differential Revision: D75977689

fbshipit-source-id: 60668af1c63fdb7240ede0736a59bd842a76de9b

Author: SamuelGabriel