Deprecate / rewrite esoteric Ops

[ricardoV94]

Description

Make our codebase more leightweight by removing:

• pytensor.tensor.basic.PermuteRowElements
• pytensor.tensor.io.LoadFromDisk (done in Remove tensor/io.py #1766)

Replace by symbolic equivalent:

• Softmax/LogSoftmax/SoftmaxGrad (use OpFromGraph to overload Softmax L_op)
• pytensor.tensor.extra_ops.Bartlett (can be implemented symbolic easily: https://github.com/numpy/numpy/blob/e7a123b2d3eca9897843791dd698c1803d9a39c2/numpy/lib/_function_base_impl.py#L3254-L3262)
• pytensor.tensor.extra_ops.UnravelIndex see Handle F-order and arbitrary index ndim in numba UnravelIndex and RavelMultiIndex #1770
• pytensor.tensor.extra_ops.RavelMultiIndex see Handle F-order and arbitrary index ndim in numba UnravelIndex and RavelMultiIndex #1770
• Several RVs: Port SymbolicRandomVariable Op to PyTensor and get rid of many RVs in PyTensor #1221
• Several ScalarOps need not exist: Remove Iv core Op in favor of Ive and add rewrite rule for log of it #1092
  • Also: xlogx, xlogy0, can just be switch statements?

[ricardoV94]

Here is the code to get a list of all Ops (with some base classes in the mix):

from pytensor.graph.op import Op

def get_all_subclasses(cls):
    all_subclasses = []

    for subclass in cls.__subclasses__():
        all_subclasses.append(subclass)
        all_subclasses.extend(get_all_subclasses(subclass))

    return all_subclasses

for op in sorted(set(get_all_subclasses(Op)), key=str):
    print(op)