Add device bridge support for HIP and CUDA.

[stellaraccident]

• Adds a custom IREE "_test_add" op which exercises actual code generation.
• Reworks the device management layer to:
  • Initialize our Device wrapper to maintain the correspondence between IREE HalDevice and PyTorch device.
  • Annotated Device with dlpack_device_type_code.
  • Detects whether the Torch device is for real-CUDA or AMDGPU/HIP CUDA and interfaces to the outside world correctly based on this.
  • Auto-detects the AMDGPU chip and the CUDA SM version and arranges for the JIT compiler to use that.
  • Dynamically enables custom kernel registration for CUDA if it is available.

Limitations (for now):

• IREE's dlpack interop only supports contiguous tensors. Lifting this requires further interfacing with the compiler to specialize on different strided layouts.
• Device synchronization is using dlpack's implicit default-stream synchronization. Since we are currently still JIT compiling in synchronous mode, we are skating by on this. Some additional hooks and APIs are needed to properly place stream events to do it for real.
• ROCM builds of torch seem to not have an easy way to map a device back to its UUID, and I didn't have a CUDA device handy to test the ways this is done on CUDA, so both are just relying on enumeration order on multi device systems. I kept the correspondence to a single place in the code and think I know how to fix this, but it will require some poking.

Depends on an IREE bump that includes: iree-org/iree#17131 (but can go in without it since that only unblocks HIP/CUDA, which were not supported yet anyway)