Build CUDA packages on CPU agent, test on GPU agent

We're currently building on
https://github.com/conda-forge/pytorch-cpu-feedstock/blob/4d13808128b46f28380791a0e08e0a09c72dbd07/recipe/conda_build_config.yaml#L12-L13

There are not many physical GPUs available, so this creates a big bottleneck. It would be nice to build the package on a CPU agent (the vast majority of the time for a given run), and then use a GPU-enabled agent only for the testing part.

This idea has already been on the table for a long time, but needs support (at least) in smithy: https://github.com/conda-forge/conda-smithy/issues/1472.

Opening for visibility, since pytorch is most affected currently by this bottleneck (AFAIU, there's also still a physically broken & offline GPU, which further exacerbates the problem).