Use batch first always

[jostosh]

• All layers now have assume batch-first inputs and produce batch-first outputs. This impedes performance slightly for DenseSum, but offers easier integration with the rest of the tf.keras framework
• Removed DimensionPermutation as it is no longer needed
• Simplified API for PermuteAndPadScopes. Infer number of decompositions/scopes through the layer's inputs
• Small fixes for get_config in some initializers
• Bump tensorflow dep to 2.2
• Update models in libspn_keras.models: add unsupervised flag to allow tensorflow.data.Dataset with only x (and no labels), also add optional infer_no_evidence to perform completion-by-posterior-marginal for unsupervised problems
• Change name of NegativeLogMarginal to NegativeLogLikelihood
• Update examples
• Add EM learning for scale parameter of NormalLeaf
• Remove BernoulliCondition
• Remove NormalizationAxes and rename ZScoreNormalization to NormalizeStandardScore, only support sample-wise normalization
• Increment version to 0.3