Handle variable length trials and data with NaN values

[ahwillia]

Trial lengths can be variable in some experiments. We should extend twPCA to handle this by simply ignoring NaN values in the data tensor. Users still specify the data as a trials x time x neuron array, but indicate missing data with nans.

Another possibility would be to pass in a vector holding the indices of trial start or end:

model = TWPCA(n_components).fit(data, trial_start=..., trial_end=...)

[poolio]

Things that need to get fixed:

• Low-rank factor initialization. Need to replace SVD with matrix completion or other strategy.
• Initial time warps (linear scaling to same length? still use shifts?)
• Mask loss function to only consider valid timepoints

[poolio]

Fixed in #6