Version 1.2.0

Several changes to default FIt-SNE settings to make it more suitable for embedding large datasets. See this recent paper by Dmitry Kobak and Philipp Berens for more details.

Major changes to default values:
-Learning rate increased from the fixed value of 200 to max(200, N/early_exag_coeff).
-Iteration number decreased from 1000 to 750.
-Initialization is set to PCA (computed via fast SVD implementations like ARPACK).

Minor changes:
-Late exaggeration start is set to the end of early exaggeration (if late exaggeration coefficient is provided).
-Limiting max step size to 5 (solves problem encountered when learning rate set too high and attractive forces cause a small number of points to overshoot)