Merge in short motif code

[LonnekeScheffer]

• Added short motif code: MotifEncoder, a sequence dataset encoder where each feature is a motif (position-specific, possibly gapped), 1 denotes the motif is present and 0 absent. Motifs are filtered by precision and recall thresholds. Several related reports are included (to calibrate recall threshold or analyse properties of learned motifs).
• Added SimilarToPositiveSequenceEncoder, a baseline method with a single feature. The encoder memorises all positive sequences in the training set, and the feature is 1 if a sequence is within a given hamming distance from ANY positive sequence in the training set, and 0 otherwise.
• Added BinaryFeatureClassifier which takes a binary encoder (such as MotifEncoder or SimilarToPositiveSequenceEncoder) and classifies an example as positive if it has any positive features. For SimilarToPositiveSequenceEncoder there is anyways just 1 feature. For MotifEncoder, optionally a subset of complementary motifs can be learned which together can be used to classify the sequence dataset.
• Added KerasSequenceCNN which is the Mason CNN for sequence datasets.
• Also contains example weighting code, to up-weight or down-weight individual examples. This is by default supported by most ML libraries like scikit-learn, pytorch. While this code is currently not actively in use, it does work and we may use it later. For now I removed the weighting strategy that I initially used for short motif (it only made sense for mutagenesis data), but kept that code on a separate branch. This PR now only contains the 'predefined weighting' strategy where example weights are supplied in a file. Other weighting strategies can be added similarly to encodings, ML methods, etc.
• Updated documentation: installation (optional dependency: Mason CNN) and added 'motif recovery' tutorial using the short motif method.
• Update version to 2.3.0