test: mul_mat tests with huge batch size

[jeffbolznv]

tests for #19471.
vulkan fix is in #19509.

[jeffbolznv]

@reeselevine can you address the webgpu failure? @JohannesGaessler or @am17an can you address the cuda failure? For context, in #19471 with a larger -ub parameter the total batch size was more than 64k and overflowed the max workgroup count in the y or z dimension (in the vulkan backend).

[reeselevine]

#19535 should fix the WebGPU failures

[am17an]

Are these only for the F16 data-type? For large batch sizes the CUDA code falls back to using cuBLAS, I think that should be a relatively simple change vs doing for quantized data types

[jeffbolznv]

In the failing model, everything was GGML_TYPE_F32. The GGML_TYPE_F16 came from me copy/pasting another test case. We could add both if there's an interesting difference in the code paths.

[am17an]

as long as it's F16, BF16 or F32 I think #19538 will fix it (passes these tests)