Closes #709

This only applies to reduction `Op`s (e.g. `CAReduce`).

The rewrite would sometimes return a new graph identical to the original,
resulting in divergence.

This means that an `aesara.function`-compiled result, e.g. `func`, will expose the underlying
JITed function as `func.fn.jit_fn`.

This should retain most backward compatibility.

`TensorType` now supports full shape information.  The field
`TensorType.broadcastable` has been replaced by `TensorType.shape`, which
corresponds to the available static shape information.  `None`s are used to
encode unknown/no static shape information for a dimension.

These changes enforce a strict "narrowing"-only conversion policy; i.e. `Type`s
can be converted to equal or more specific types.

After fixing `Shape`'s broadcastable information,
`local_subtensor_remove_broadcastable_index` started replacing all forms of
`shape(x)[0]` with `shape(x).dimshuffle(())` when `shape(x).broadcastable
== (True,)`, so a lot of the changes in this commit compensate for that
difference.