fix: performance of ufunc resolution for non-nominal signatures

[agoose77]

@lgray observed that find_ufunc is a significant bottleneck in coffea-dask workflows. Whilst we can optimise this further by changing our internal representation of behaviors (i.e. __getitem__ registers overloads in high-performance lookup structures), for now I suspect that we can improve the performance of the common-path. @jpivarski notes that the behaviour changed in this PR was never intended to be supported, so it represents a correctness improvement regardless.

The case that this PR improves is

np.add(foo, bar, ...) where neither foo nor bar are Awkward Arrays with names (e.g. __record__). This represents most ufunc operations given that even ufunc overloads of records often delegate to Awkwards' ufunc mechanism on arrays of numbers.

[lgray]

This corresponds to a ~20% improvement in execution time (O(1s) out of 5s) building a single task graph in the workflow in question, so this is a rather significant improvement.

Keeping in mind the task graph gets built 180 times or so for the full analysis.

[jpivarski]

I think this is ready to be merged. I've check-marked that all of the rules I'd expect are included in the new implementation.

Thanks!