Deeplift

Corresponds to PR https://github.com/kundajelab/deeplift/pull/117 (sigmoid as the output layer was already supported)

Thanks to atseng95 for PR https://github.com/kundajelab/deeplift/pull/109, which greatly speeds up dinucleotide shuffling. This tag also incorporates the change from PR https://github.com/kundajelab/deeplift/pull/105 (which has support for supplying pre-generated shuffled references, and was tagged as v0.6.11.0), as well as the small fix in PR https://github.com/kundajelab/deeplift/pull/101 (which allows the user to recover if they accidentally set an invalid task index, and was tagged as v0.6.10.1)

Corresponds to the fix in https://github.com/kundajelab/deeplift/pull/96 by berleon, who found that it greatly improves the compilation time for models that have layers with multiple inputs.

Corresponds to features implemented in PR https://github.com/kundajelab/deeplift/pull/93. Two features: (1) has a fix for loading models that don't have biases, and (2) a message about the target layer that would previously be thrown as a runtime error now just results in a warning message being printed, as there are legitimate situations where that edge case can occur. See https://github.com/kundajelab/deeplift/issues/92 for the issue that prompted the changes.

This feature was requested in https://github.com/kundajelab/deeplift/issues/83 and was implemented in PR https://github.com/kundajelab/deeplift/pull/84/files. I forgot to merge it into the master branch at the time and am doing so now. The genomics notebook was updated to use this feature (and also updated to python 3) in https://github.com/kundajelab/deeplift/pull/94

(so that the random state doesn't have to be controlled by setting the numpy random seed externally, which doesn't always play will with jupyter notebooks)