[RFC] Support for __array__()

[gmarkall]

Pushing this up early to request comments / feedback - I think this is one way to solve rapidsai/cudf#15694, with broad applicability to other use cases - this should allow Numba-jitted functions to be called on anything that has an __array__() function, much like CUDA kernels can be called on anything that supports the CUDA Array Interface.

The implementation handles typing and unboxing by falling back to using __array__() only if treating the object as an ndarray fails - therefore, it should not impact the fastest path through dispatch - the code additions here are only executed in cases where the dispatch would otherwise have failed.

One minimal test case is added to demonstrate the idea. Additional tests I think would be needed:

• Testing with and without NRT enabled, though I'm not sure we still have a way to disable NRT. If we can't disable NRT anymore, then there's some dead code like numba_adapt_ndarray that can be deleted.
• Testing with a mix of __array__() and ndarray arguments.
• Testing that this doesn't cause any errors in reference counting / memory leaks (not that I expect it should, but having a test to prove it would be nice)

cc @brandon-b-miller @AjayThorve

Another benefit of this PR - the following:

import pandas as pd
from numba import njit

s1 = pd.Series([1, 2, 3])
s2 = pd.Series([4, 5, 6])

df = pd.DataFrame([s1, s2])

@njit
def f(x):
    return x + 1

print(f(s1))
print(f(df))

presently fails on main with an error ending in:

This error may have been caused by the following argument(s):
- argument 0: Cannot determine Numba type of <class 'pandas.core.series.Series'>

but with this PR, gives:

$ python repro.py 
[2 3 4]
[[2 3 4]
 [5 6 7]]

In general, it becomes possible to pass pandas objects directly to jitted functions.

[gmarkall]

Also - I can't find a feature request issue to link this too - I'm surprised if this hasn't been asked-for directly at some point though 🤷

[gmarkall]

The CI failure is triggered by the latent bug #9585.

[gmarkall]

I think the CI failure calls to attention that I need an additional try: ... except ValueError: ... around this, but that wouldn't be sufficient to make this work, due to #9585 - the NotImplementedError it produces would not be caught by except ValueError.

[gmarkall]

Notes from triage meeting (very quickly written, will tidy up later):

• This should make it possible to pass a pandas series or dask array to a jitted function
• If there's a performance issue due to the use of array function, (e.g. CuPY array converting to NumPy array), then the library doing the inefficient thing should be responsible for any warnings
• Would we want a warning to be emitted when this is done?