fix: replace protocol with direct subclass

[agoose77]

Python typing protocols make it possible to enforce duck-typing at type-check time, i.e. ensure that an object has the appropriate methods and attributes rather than whether it inherits from another class. To partially check this at runtime there is a runtime_checkable decorator that simply test whether a given object implements the same methods (ignoring their types).

However, this apparently is much slower than I'd anticipated #2028. This issue is only a problem because we need to sanitise the user inputs to our layouts. Otherwise, once mypy is fully covered through the code-base, we could just remove our runtime input validation in most cases. Given that we want to provide helpful user errors, we need runtime checks to work. We could fast-path this with

def isinstance_(obj, type_) -> bool:
    return type.__instancecheck__(type_, obj) or type_.__instancecheck__(obj)  

but in this case, we don't truly need to duck-type backends; they're a limited abstraction that we can forgo the flexibility of protocols.

Fixes #2028

[codecov]

Codecov Report

Merging #2029 (61f2903) into main (a242c6e) will decrease coverage by 0.00%.
The diff coverage is 100.00%.

❗ Current head 61f2903 differs from pull request most recent head 3e366f1. Consider uploading reports for the commit 3e366f1 to get more accurate results

Additional details and impacted files

Impacted Files	Coverage Δ
src/awkward/_backends.py	84.40% <100.00%> (+0.19%)	⬆️

[jpivarski]

I appreciate what the typing is expressing about constraints here, but as @nsmith- showed, it's adding too much time.

We talked on Zoom about this. Since we are requiring that the backend is one of (our own) four backend objects, the test could be a function in ak._util:

def is_backend(backend):
    return (
        backend is ak._backends.NumpyBackend
        or backend is ak._backends.CupyBackend
        or backend is ak._backends.JaxBackend
        or backend is ak._backends.TypeTracerBackend
    )

which is a centralized place to update if we ever add a fifth backend. That could be even faster than a standard isinstance check. (I'm describing an is check on each possibility, rather than a backend in (...) to even avoid __eq__ checks.)

But that's probably not necessary; this PR solves the immediate performance issue as-is.