Added options for `disconnected_outputs` to Rop.

Fix logic for timing / choosing algo in dnn_fwd

Implement Grouped Convolutions

pr mac tests

Flake8 for test_blas.py

Use GpuDnnReduction to replace GpuMaxAndArgmax when possible.

Fix alerts found with lgtm

Only compare to scipy timing in FAST_RUN.

Make sure MRG uniform in float16 do not return 0

Implement grad for AbstractBatchNormTrainGrad

Add debug code to print algo chosen at runtime for cuDNN convolutions.