GitLab

* Guard the unsafe tf.exp to prevent Inf cost

* Update run_lfads.py

Add the hyperparameter `_clip_value` to resolve the issue in the prototype example

For stitched models, the readin matrices and bias vectors are initialized to the
"alignment" matrix and bias specified in each dataset's .h5 file. If do_train_readin is
True, these will be trainable, and if not, they will be fixed.

Made PCA bias work for input alignment. Fixed output matrix to pinv, added gaussian example, fixed output gaussian param inits, updated README