Replay-classification

+ Fix the state transition problem from the last release (v0.6.0) by making the user estimate the movement standard deviation.

+ Now uses scikit-learn for density estimation
+ Implement smoother to use both past and future information
+ Fix some integration errors by including bin size
+ Use tqdm to report fit progress
+ Allow bins, bin size or number of bins to be specified by user
+ Reorganized Decoders to use a base class
+ Change spike data dimensions
+ Handle NaNs better

+ Fix issue where predicting on values outside of support in clusterless decoding results in *inf* values.

+ Handle NaN coefficients in sorted spike decoder from GLM. This can happen when there are no spikes in the training data.

+ Require user to input lagged position (that is the position from the previous time step) because if you filter the position before input (e.g. you wanted to keep only correct trials), then it is impossible to figure out the position from the previous time step (cc18e46)
+ Allow user to specify sampling frequency in the sorted spike observation model plot so that the firing rate is correct (cdb129e)
+ Allow user to specify position knot spacing in sorted spikes model instead of placing knots at quantiles (255f7c4)
+ Use `regularized_glm` package for fitting sorted spike models (0be8b8b) in order to speed up the fitting of the L2-penalized observation model (0be8b8b).
+ Reorganize so that fit functions are in their own modules (9bf7a84).

+ Fixed error when design matrix contained NaNs
+ Fixed tests that broke in previous release