test-conv1d and test-conv2d failed on GPUs with computation capability <= 6.1

[bssrdf]

Just to report right now two test cases test-conv1d and test-conv2d cannot pass if running on GPU backend and its computation capability <= 6.1. On a GTX 1070 with CUDA version 12.1, it gave
ggml_im2col (4320): PASSED
ggml_conv2d (480): FAILED

This might be related to this section of the code

ggml/src/ggml-cuda.cu

Lines 7618 to 7639 in 3d57e76

	else {
	cuda_pool_alloc<float> src0_ddq_as_f32;
	if (src0->type != GGML_TYPE_F32) {
	const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(src0->type);
	GGML_ASSERT(to_fp32_cuda != nullptr);
	src0_ddq_as_f32.alloc(row_diff*ne00);
	to_fp32_cuda(src0_dd_i, src0_ddq_as_f32.get(), row_diff*ne00, stream);
	}
	const float * src0_ddf_i = src0->type == GGML_TYPE_F32 ? (const float *) src0_dd_i : src0_ddq_as_f32.get();
	const float alpha = 1.0f;
	const float beta = 0.0f;
	CUBLAS_CHECK(cublasSetStream(g_cublas_handles[id], stream));
	CUBLAS_CHECK(
	cublasSgemm(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N,
	row_diff, src1_ncols, ne10,
	&alpha, src0_ddf_i, ne00,
	src1_ddf_i, ne10,
	&beta, dst_dd_i, ldc));
	}

Here src1 is not converted to FP32 while src0 is.