Add Cooperative Groups API integration

[thedodd]

This works as follows:

• Users build their Cuda code via CudaBuilder as normal.
• If they want to use the cooperative groups API, then in their build.rs, just after building their PTX, they will:
  • Create a cuda_builder::cg::CooperativeGroups instance,
  • Add any needed opts for building the Cooperative Groups API bridge code (-arch=sm_* and so on),
  • Add their newly built PTX code to be linked with the CG API, which can include multiple PTX, cubin or fatbin files,
  • Call .compile(..), which will spit out a fully linked cubin,
• In the user's main application code, instead of using launch! to schedule their GPU work, they will now use launch_cooperative!.

todo

• Update cust to expose the cuLaunchCooperativeKernel in a nice interface. We can add the cooperative multi device bits later, along with all of the other bits from the cooperative API.
• Update cuda_builder crate to expose a nice API around linking PTX code with the Cooperative Groups API bridge code.
• Remove the Justfile. I was only using it for POC testing.
• Remove the PTX code from this branch, currently it is just here for reference.

---

• closes Add Cooperative Groups API integration #87