ENH: Add __slots__ to private (sub-)classes in numpy.lib._index_tricks_impl

[jorenham]

This adds __slots__ to the (private) classes in numpy.lib._index_tricks_impl and their subclasses, improving the performance of:

• numpy.mgrid
• numpy.ogrid
• numpy.r_
• numpy.c_
• numpy.index_exp
• numpy.s_
• numpy.ma.mr_

Note

I initially wanted to add __slots__ to ndenumerate and ndindex as well, but decided not to, as there's a (small) chance that it will be backwards compatible.:
Because they are part of the public API, and not marked as @final, it's possible that > they're used as base-class for user-defined types. And while that, on its own, is not a problem here, it could become one if such a user-defined type has more than one base class.

But if this isn't a problem, I'd be happy to add __slots__ to these classes, too :)

[eendebakpt]

@jorenham Do you have some benchmarks for the methods affected by these changes? I suspect this will mostly impact operations involving small arrays

[jorenham]

@eendebakpt A quick %timeit comparison on Python 3.12 gave the following results:

%timeit	numpy/numpy@main	jorenham/numpy@index_tricks-__slots__	Speedup
np.mgrid[0:2]	3.74 μs ± 10.1 ns	3.18 μs ± 7.03 ns	+ 17.6 %
np.mgrid[0:1000]	4.78 μs ± 7.69 ns	4.12 μs ± 10.4 ns	+ 16.0 %
np.mgrid[0:2, 0:2]	23.1 μs ± 49.3 ns	20.4 μs ± 43.3 ns	+ 14.9 %
np.r_[0, 1]	6.65 μs ± 9.88 ns	5.61 μs ± 8.72 ns	+ 18.5 %
np.r_[0, 1, 2, 3]	8.34 μs ± 8.4 ns	7.15 μs ± 15 ns	+ 16.6 %
np.r_[[[0]], [[1]]]	20.1 μs ± 31 ns	17.5 μs ± 30.3 ns	+ 14.9 %
np.s_[2::2]	95.4 ns ± 0.0396 ns	45.7 ns ± 0.0559 ns	+ 108 %

The remaining ogrid, c_/ma.mr_, and index_exp should have similar results as mgrid, r_, and s_, respectively.

[eendebakpt]

Thanks for the benchmarks, looks like a nice improvement with no regressions.

Normally we leave whitespace fixes out of the PR to make reviews easier, but it is ok to keep them in afaic.

[mattip]

Thanks @jorenham