[AMD] refactor proton to use rocprofiler-sdk and deprecate roctracer

[ZelboK]

1. tests seem to pass locally running

export ROCP_TOOL_LIBRARIES=$(python -c "import triton._C.libproton as p; print(p.__file__)")
export LD_PRELOAD=/opt/rocm/lib/librocprofiler-sdk-roctx.so # details on why needed later
pytest third_party/proton/test/test_profile.py

I am not comfortable enough with the CI to know best way for figuring out how to incorporate this for these tests.

2. Right now lazy loading isn't supported by rocprofiler-sdk. As a result, since we need to initialize it before hip does to properly intercept kernels, I am using ROCP_TOOL_LIBRARIES to get around that. PyTorch will initialize hip if we don't do it first. rocprofiler-sdk looks for this env variable as i understand it and then calls rocprofiler_configure(note for self: audit later).

3. rocprofiler-sdk most importantly supports the ability to attach and detach at any point so you can minimize overhead by not having to profile all the time.

4. proton-cli is the way to use the new library, as a consequence of 2. Lazy loading may be added in the future.

5. We use rocprofiler-register to make our lives easier. Each library (HIP, ROCr) calls into the register and provide interception table. That's the mechanism widely accepted (as a standard) in ROCm stack.

6. Pytorch seems to be using this /opt/rocm/lib/libroctx64.so but rocprofiler-sdk intercepts /opt/rocm/lib/librocprofiler-sdk-roctx.so. A few of the nvtx tests were failing

For this PR I think the ergonomics of the libraries we're loading will need to be looked at to make sure they're ideal.

NOTE: Stochastic sampling will be added in a future PR. The work is complete but it seems to work.