Precomputing batch size in MB

[BalzaniEdoardo]

Aim to facilitate batch size selection for stochastic gradient method when fitting very large dataset. Especially useful as a lower bound when estimating the GPU resources necessary by model fitting.

This method will be part of Basis, and can use the num_output_features parameter to estimate the size of the output array that the basis generates, given the number of samples in a batch and, optionally, the dtype of the array.