Support array_like low/high bounds for randint

[vlad-perevezentsev]

This PR adds array_like low/high support to randint to align with numpy.random.RandomState.randint

It adds new C routines irk_rand_<type>_broadcast for all 9 integer dtypes, drawing per-element integers over [low, high) with Lemire's multiply-shift (the same method as NumPy).
An earlier masked-rejection version benchmarked slower so it was replaced.

A mulhi helper returns the high half of the product because Lemire computes word × range which overflows the type so the 64-bit case uses a 32-bit schoolbook multiply.

It also groups all randint tests into TestRandint class and adds new ones.

[ndgrigorian]

@vlad-perevezentsev
We will want to apply to randint_untyped as well

[ndgrigorian]

@vlad-perevezentsev
Edge case issue:

In [5]: np.random.randint([3], [9], size=())
Out[5]: array(6)

In [6]: rand.randint([3], [9], size=())
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
Cell In[6], line 1
----> 1 rand.randint([3], [9], size=())

File mkl_random/mklrand.pyx:2455, in mkl_random.mklrand._MKLRandomState.randint()
-> 2455 'Could not get source, probably due dynamically evaluated source code.'

File mkl_random/mklrand.pyx:2302, in mkl_random.mklrand._MKLRandomState._randint_broadcast()
-> 2302 'Could not get source, probably due dynamically evaluated source code.'

File ~/miniforge3/envs/dpctl_dev/lib/python3.14t/site-packages/numpy/lib/_stride_tricks_impl.py:443, in broadcast_to(array, shape, subok)
    400 @array_function_dispatch(_broadcast_to_dispatcher, module='numpy')
    401 def broadcast_to(array, shape, subok=False):
    402     """Broadcast an array to a new shape.
    403
    404     Parameters
   (...)    441            [1, 2, 3]])
    442     """
--> 443     return _broadcast_to(array, shape, subok=subok, readonly=True)

File ~/miniforge3/envs/dpctl_dev/lib/python3.14t/site-packages/numpy/lib/_stride_tricks_impl.py:377, in _broadcast_to(array, shape, subok, readonly)
    375 array = np.array(array, copy=None, subok=subok)
    376 if not shape and array.shape:
--> 377     raise ValueError('cannot broadcast a non-scalar to a scalar array')
    378 if any(size < 0 for size in shape):
    379     raise ValueError('all elements of broadcast shape must be non-'
    380                      'negative')

ValueError: cannot broadcast a non-scalar to a scalar array

[ndgrigorian]

@vlad-perevezentsev
We get incorrect results (OOB) for smaller int types:

In [1]: import mkl_random as rand
<frozen importlib._bootstrap>:491: RuntimeWarning: The global interpreter lock (GIL) has been enabled to load module 'mkl_random.mklrand', which has not declared that it can run safely without the GIL. To override this behavior and keep the GIL disabled (at your own risk), run with PYTHON_GIL=0 or -Xgil=0.

In [2]: import numpy as np

In [3]: rand.randint([-5], [6], dtype=np.int8)
Out[3]: array([-119], dtype=int8)

In [4]: np.random.randint([-5], [6], dtype=np.int8)
Out[4]: array([-5], dtype=int8)

[ndgrigorian]

I think it would be good at some point to experiment with chunking these loops rather than going with a single-pass. Can we add a TODO comment to remind of experimenting?

[ndgrigorian]

This LGTM, in the follow-up we should fix the remaining types and also fix randint_untyped to align with the others