Synthetic data generation

[berkgercek]

Some functionality for synthetic data generation already exists in the prior localization repository in synth.py and glm_predict.py.

These should be reorganized in a sensible way to:

1. Generate predicted spike counts in bins from a new design matrix with the same covariates, or a subset of the existing design matrix, using a nglm.predict() method.

2. Decompose PETHs into their subcomponents contributed by each kernel

3. Optionally generate totally synthetic spike times, which is harder especially if you want to randomize when the spike occurs within a bin. This could be useful for testing other data analysis methods or pipelines.

4. and 2. are high-priority and low-hanging fruit given that the primitives for 1. exist in sklearn and the code for 2. has already been written. They just need to be adapted.