Use dpctl.tensor.matmul in the backend of dpnp.matmul when inputs are integer

[vtavana]

resolves #2270

OneMath (OneMKL) routines (gemm, gemv, gemm_batch) for matrix multiplication only support floating point data types. If inputs are integer, to use OneMath we need to upcasting them to floating point dtypes, perform the calculation and then cast back the result to integer dtypes which is unsafe and we may loose some information for large integers.
In this PR, the logic for dpnp.matmul is updated to use dpctl.tensor.matmul when result has a integer dtypes.

Performance Analysis

$ sycl-ls
[level_zero:gpu][level_zero:0] Intel(R) oneAPI Unified Runtime over Level-Zero, Intel(R) Data Center GPU Max 1100 12.60.7 [1.6.31294+9]
[opencl:cpu][opencl:0] Intel(R) OpenCL, Intel(R) Xeon(R) Platinum 8480+ OpenCL 3.0 (Build 0) [2025.19.1.0.16_160000]
[opencl:gpu][opencl:1] Intel(R) OpenCL Graphics, Intel(R) Data Center GPU Max 1100 OpenCL 3.0 NEO  [24.39.31294]

import numpy, dpnp

# Case 1 - matrix-matrix multiplication
n=1024
a=numpy.ones((n,n), dtype="i8")
%timeit numpy.matmul(a, a)

ia=dpnp.array(a, device="gpu")
%timeit dpnp.matmul(ia, ia); ia.sycl_queue.wait()

# Case 2 - Matrix-vector multiplication
n=4096*4
a=numpy.ones((n,n), dtype="i8")
b=numpy.ones((n,), dtype="i8")
%timeit numpy.matmul(a, b)

ia=dpnp.array(a, device="gpu")
ib=dpnp.array(b, device="gpu")
%timeit dpnp.matmul(ia, ib); ia.sycl_queue.wait()

# Case 3 - Vector-matrix multiplication
n=4096*4
a=numpy.ones((n,n), dtype="i8")
b=numpy.ones((n,), dtype="i8")
%timeit numpy.matmul(b, a)

ia=dpnp.array(a, device="gpu")
ib=dpnp.array(b, device="gpu")
%timeit dpnp.matmul(ib, ia); ia.sycl_queue.wait()

# Case 4 - batch matrix-matrix multiplication
n=256
a=numpy.ones((n,n,n), dtype="i8")
%timeit numpy.matmul(a, a)

ia=dpnp.array(a, device="gpu")
%timeit dpnp.matmul(ia, ia); ia.sycl_queue.wait()

Case # - n	NumPy	dpnp - Xeon	dpnp - PVC
Case1 - 1024	4.98 s ± 867 μs	9.82 ms ± 40.6 μs	3.34 ms ± 4.16 μs
Case2 - 4096×16	4.2 s ± 616 μs	9.77 s ± 112 ms	1.27 s ± 2.04 ms
Case3 - 4096×4	9.6 s ± 3.3 ms	94.6 ms ± 2.15 ms	50.9 ms ± 38.2 μs
Case4 - 256	3.19 s ± 291 μs	48.9 ms ± 244 μs	10.4 ms ± 32.3 μs

For all cases dpnp shows a better performance compared to numpy except for Case 2 on Xeon.

• Have you provided a meaningful PR description?
• Have you added a test, reproducer or referred to issue with a reproducer?
• Have you tested your changes locally for CPU and GPU devices?
• Have you made sure that new changes do not introduce compiler warnings?
• Have you checked performance impact of proposed changes?
• If this PR is a work in progress, are you filing the PR as a draft?

[github-actions]

View rendered docs @ https://intelpython.github.io/dpnp/index.html

[coveralls]

coverage: 71.629% (+0.06%) from 71.572%
when pulling 1edddbe on fix_issue-2270
into 0c455a6 on master.

[github-actions]

Array API standard conformance tests for dpnp=0.17.0dev5=py312he4f9c94_31 ran successfully.
Passed: 971
Failed: 0
Skipped: 29

[antonwolfy]

Thank you @vtavana , I leave two more minor comments, but overall LGTM!