cuda : add FILL op support

[JayZenith]

Add CUDA backend support for the GGML_OP_FILL, which was previously missing (CPU and Vulkan had it). This operation is used by the Qwen3-Next model (discussion in #16623).

• Added fill.cu and fill.cuh with a simple CUDA kernel.
• Added dispatch case in ggml_cuda_compute_forward()
• Declared support in ggml_backend_cuda_device_supports_op()

Tested with test-backend-ops -o FILL on Tesla T4:
FILL(type=f32,ne=[10,10,4,3],c=0.000000): OK
FILL(type=f32,ne=[303,207,11,3],c=2.000000): OK
FILL(type=f32,ne=[800,600,4,4],c=-152.000000): OK
FILL(type=f32,ne=[2048,512,2,2],c=3.500000): OK
4/4 tests passed

[am17an]

I'm not sure why this should just not be a cudaMemsetAsync

[JayZenith]

@am17an cudaMemsetAsync writes only one byte value repeatedly and dosent interpret floats/doubles. It works for 0.0f but fails for numbers like 1.0f (0x3F800000) as it would write 0x3F to every byte. This kernel writes the full float/double per element, so works for any number. Essentially, byte-wise vs element-wise writing.

[am17an]

You need to also enable this kernel via ggml_backend_cuda_device_supports_op, right now I'm not sure how it's passing test-backend-ops for you