fix thread queue allocation on windows

[mattip]

This was worked through on MacPython/openblas-libs#284 to fix numpy/numpy#27989 and MacPython/openblas-libs#283.

On machines with many cores, there was cross-thread use of the wrong queues. This is solved here.

I used Claude code to identify and fix the problem.

Maybe related to #1404

[mattip]

I would add the issue reproducer as a test, but it needs a machine with more cores than is available on CI, especially since the OpenBLAS default is 64 threads on windows (openblas-libs uses 24).

[martin-frbg]

Thanks - would it be possible to attach the reproducer here (though the biggest Windows machine I have myself is only 10 cores/20 threads) ?

Not sure what was up with issue 1404 - I'm pretty sure the Windows support in OpenBLAS was even worse back then and compilers have also improved. Anyway the cpu mentioned there had only 6 cores/12 threads. Guess I'll finally close #1404 as not reproducible anymore (after doing one final build with Haswell rather than CooperLake target)

[mattip]

Here is the reproducer

import numpy as np
from concurrent.futures import ThreadPoolExecutor

def crash_test(size_a):
    a = np.random.random((size_a,4))
    b = np.random.random((4,4))
    c = np.matmul(a,b)
    return c.mean()
    
num_threads = 4
counts = 128
futures = []
total = 0
size_a = 1024*1024

with ThreadPoolExecutor(max_workers=num_threads) as executor:
    for i in range(counts):
        futures.append(executor.submit(crash_test, size_a))
    for i in range(counts):    
        total += futures[i].result()
print(total)