N2v

* Fix number of blind-spot computation: 0.198% of the pixels are manipulated per patch.
* Pin tensorflow and keras version numbers in README instructions
* Fix 42: Tiling during prediction keeps number of axes consistent with the input.
* Fix 45: Extract n patches from each sample S.
* Fix 43: Images are cast to 32bit float type during prediction.
* Add augmentation flat to training script entry point.
* Add BSD68 example notebook. This notebook reproduces the results reported in the paper (CVPR'19).

* Add optional patch shuffling parameter
* Add progress bar to validation data preparation
* Normalize each channel separately
* Increase keras version to 2.2.5
* Relax csbdeep version requirement from 0.4.0 to 0.4.0<=csbdeep<0.5.0

* Increase CSBDeep version from 0.3.0 to 0.4.0

This release includes:
* The data generator now sorts the filenames before images are loaded.
* This release includes entry points of the trainN2V.py and predict n2v.py scripts.