Run CI on all Windows Driver models

I don't like bugs that stem from insufficient testing. This https://github.com/NVIDIA/cccl/issues/10716 is one of them.

Our Windows CI only runs the tests on `WDDM` mode. There are two other driver modes `TCC|MCDM` that users might use CCCL in, when on Windows. Code might fail on one mode while run on another.

**Should we add TCC/MCDM testing in our CI?**

cuda-python folk have a testing matrix that spans across all three. My discussion with @wmaxey showed that this might not be trivial.

- We run our tests in containers (cuda-python doesn't) - and a TCC GPU may not be reachable from inside a Windows container at all.
- So we'd likely have to run the test job on the bare host, outside the devcontainer environment and then install VS Build Tools + CUDA toolkit manually (or ship prebuilt test binaries to the runner as artifacts)
- Flip the driver to TCC
- Run the tests