This helper could arbitrarily override the default output_type from `ScalarOp.make_node` so that the output type matched one of the input types.
This can be used to create artificial Op signatures that don't make sense or can't be cleanly implemented in other backends. For instance an Add with signature (int8,int64)->int8.

This helper was historically used in:
 1. Elemwise inplace rewrite. I assume as a preventive measure. However, regular use should never require changing the ScalarOp signature, as we only try to inplace on inputs that match the output dtype and recreating the same Op with the same input types should always return the same output type. ScalarOp don't have a concept of inplace, only the Elemwise wrapper does, and it shouldn't require recreating/mutating the original Op.
 2. SecondOp. Here it makes sense, but a custom static_method works just as well
 3. Inplace tests with the inplace variants of `@scalar_elemwise` decorator. These test Classes were removed. It still didn't make sense to test/force Ops to have an artifical signature for the sake of tests. They were removed anyway

Note: Numba object mode fallback is not safe with lists

Closes https://github.com/pymc-devs/pytensor/issues/1245

It's more readable and avoids potential bugs when force_unique is not set to True

Implementation was specializing on node repeated inputs an `unique_names` would return the same name for repeated inputs. The cache key didn't account for this.

We also don't want to compile different functions for different patterns of repeated inputs as it doesn't translate to an obvious handle for the compiler to specialize upon. We we wanted to inline constants that may make more sense.

It's necessary to encode the edge information, not only the nodes and their ordering

* Update `run_mypy.py`

* add `annotation-unchecked`

* remove test exclusion

Add continue-on-error: true to the script step to prevent the entire
workflow from failing when the script encounters an error. This allows
the scheduled workflow to complete successfully even if the script fails.

* Force environment dependent xfail test to fail

* New version of MLX crashing in CI

This way they can be fully cached when re-running tests

Direct import is not properly mocked by tests when trying to run `compare_numba_and_py` with `eval_obj_mode=True`

These tests were covering things that don't exist anymore. params in python perform method of Ops, or misbehavior of an Op not respecting the signature