Empty DimShuffles would return a scalar instead of an array, which can then lead to errors in ops that expect an array.

It doesn't make sense to include `fast_run` if `fast_compile` mode is being used. Some rewrites such as the FusionOptimizer are not compatible with `fast_compile` mode which prevents the creation of C thunks. The FusionOptimizer has no way of knowing this is the case, and assumes it is safe to return Composites with more than 32 operands, even though that's not the case with the Python perform method.

This was not an issue in my local machine, but failed on the Github CI. It could be due to compiler optimizations. Case 69 used to look like this:

```python
 Elemwise{Composite{(i0 * tan(i0) * tan(i0) * i1)}} [id C]
 |x [id A]
 |x [id A]
 ```

 And now looks like this
 ```python
 Elemwise{Composite{(i0 * tan(i0) * tan(i0) * i0)}} [id C]
 |x [id A] [None]
 ```

* Move local_add_mul_fusion to `rewriting/elemwise` and remove unused/duplicated TestAddMulFusion tests
* Use EquilibriumGraphRewriter for local_add_mul_fusion
* Do not register optional rewrites if tensor__local_elemwise_fusion flag is disabled