CI: Make the `fetch_ctk` action take component names as input for caching purposes

[leofang]

We currently hardwire what components to download and use in the CI:

cuda-python/.github/actions/fetch_ctk/action.yml

Lines 85 to 92 in d425a88

	populate_cuda_path cuda_nvcc
	populate_cuda_path cuda_cudart
	populate_cuda_path cuda_nvrtc
	populate_cuda_path cuda_profiler_api
	populate_cuda_path cuda_cccl
	if [[ "$(cut -d '.' -f 1 <<< ${{ inputs.cuda-version }})" -ge 12 ]]; then
	populate_cuda_path libnvjitlink
	fi

The artifacts are cached in GitHub Caches space for reuse:
https://github.com/NVIDIA/cuda-python/actions/caches
However, it means that if we need to update this component list, it involves a maintainer (me 😓) using admin privilege to remove the cache first. I can do that no problem, but it is better to make it developer-configurable so that

• the action takes a list of components as inputs, same as the other 3rd-party action we use (as a bandaid!)
  

  cuda-python/.github/workflows/test-wheel-windows.yml

  Lines 161 to 165 in 0dfae43

  	uses: Jimver/cuda-toolkit@v0.2.21
  	with:
  	cuda: ${{ inputs.cuda-version }}
  	method: 'network'
  	sub-packages: ${{ env.MINI_CTK_DEPS }}

• the action computes the cache key using the component names, so that when the list is updated the cache is invalidated automatically

cc @cryos @carterbox for vis

[carterbox]

Assign me, please