[Python] If numpy is available use it for normalizing slice

[raulcd]

Describe the enhancement requested

There were some concerns after the following PR was merged on the performance about it:

• GH-46606: [Python] Do not require numpy when normalizing slice #46732

I've tested with both previous case and new case with the following results:
Previously to the PR using numpy:

In [2]: a = pa.array(np.arange(0, 2_000_000))

In [3]: %timeit a[slice(1, 1_000_000, 2)]
763 μs ± 4.28 μs per loop (mean ± std. dev. of 7 runs, 1,000 loops each)

After the PR using Python object:

In [2]: a = pa.array(np.arange(0, 2_000_000))

In [3]: %timeit a[slice(1, 1_000_000, 2)]
25.7 ms ± 703 μs per loop (mean ± std. dev. of 7 runs, 10 loops each)

We should probably use np.arange if numpy is available otherwise revert to using a Python object.

We could also potentially add the possibility for Array::Slice to have a step attribute being able to remove some of the custom code and just use step internally.

Component(s)

Python

[pitrou]

We should probably use np.arange if numpy is available otherwise revert to using a Python object.

Or, instead of exhibiting wildly different performance based on Numpy availability, perhaps we can implement a pa.arange function? It should relatively easy to do so in C++ (we can limit the output types to integer types for now).