CUDA: fix MMQ memory fault if n_expert >> n_ubatch

[JohannesGaessler]

My attention was drawn to these failing test cases by @am17an :

for (bool b : {false, true}) {
    test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_Q4_0, GGML_TYPE_F32, 512, 10, b, 640, 508, 2560));
}

With the CUDA backend they result in an illegal memory access. The issue seems to be a bug in determining how much padding is needed for the temporary q8_1 buffer in the MMQ kernel. It's supposed to be as many extra columns as is the maximum tile width J. For dense models this is determined bounded by src1->ne[1]. For MoE models however the tensor layout is different and it is instead bounded by src1->ne[2]. So the MMQ code is using the wrong value which can result in OOB memory accesses if the number of experts is much larger than the physical batch size. And the fix is simply to use the correct value instead.

I decided against adding test cases to test-backend-ops.cpp because the tensors needed for a reproduction are relatively large and the conditions for this defect to manifest as a bug were highly specific.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: No

[JohannesGaessler]

Fixes #29847 . (Seems this is where the reproduction is from which with I was pinged.)

Supersedes #27044 . The fix in that PR should work to resolve the illegal memory access but also overallocate memory.

May be the same issue as #28383 .