Fix/issue 1433

[David366AI]

Fix bug #1433

In file ops.cpp

// ggml_compute_forward_im2col_f16
// src0: kernel [OC, IC, KH, KW]
// src1: image [N, IC, IH, IW]
// dst: result [N, OH, OW, ICKHKW]
static void ggml_compute_forward_im2col_f16(
const ggml_compute_params * params,
ggml_tensor * dst) {

const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];

GGML_ASSERT(src0->type == GGML_TYPE_F16);
GGML_ASSERT(src1->type == GGML_TYPE_F32); // This line is not so reasonable.
GGML_ASSERT( dst->type == GGML_TYPE_F16);
The follow picture is from one onnx (yolo26 export half onnx).
dst is F16, src0 (Kernel) is F16, image is F16, but
GGML_ASSERT(src1->type == GGML_TYPE_F32); // Assert failed.
It doesn't make sense.

Suggest to change to be:
GGML_ASSERT(src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16);

[David366AI]

The solution only fix CPU backend.

If you approve it, I could fix cuda/vulkan/Metal etc.

[ggerganov]

@David366AI Thank you for the contributions. Waiting for the CI to pass and will merge.