rel-2.2.0

The following changes were made in this release:

• TensorConstant is now a subclass of TensorVariable
  • Code that was previously using isinstance(..., TensorVariable) as a means of filtering out TensorConstants will need to be updated
• The following config flags have been removed: gpu_*_elemwise, warn__inc_subtensor1_opt, warn__inc_set_subtensor1, warn__reduce_join, warn__signal_conv2d_interface, warn__vm_gc_bug, warn__gpu_set_subtensor1, warn__subtensor_merge_bug, warn__sum_div_dimshuffle_bug, warn__gpusum_01_011_0111_bug, warn__argmax_pushdown_bug, warn__identify_1pexp_bug.
• Fixed a bug in an aesara.scan warning message
• local_1msigmoid rewrites are applied by default
• PatternSub will no longer return nodes with different output types
• Fixed a bug in PatternSub when get_nodes and values_eq_approx are specified
• Implemented DeepCopyOp.infer_shape
• Added a CholeskySolve Op
• Fixed a scope issue in the C code generated by UltraFastScalarSigmoid