Use new UR handles for opencl instead of casting mechanism

[omarahmed1111]

This PR changes sycl-rt for opencl interop functionality to use the new UR handles layer for opencl, instead of relying on casting directly handles directly. So, it removes the 1:1 relation between UR/sycl-rt and opencl.

The changes in this PR involves:

• Changing sycl interop API to not own the opencl native handle passed to sycl-rt by default. SYCL application could be still using the native opencl handle outside of SYCL-RT.
• Remove the specialized opencl retains in sycl-rt as this will retain UR objects which are already owned by sycl-rt to make it and release it. Before this PR, this was needed as sycl-rt was being passed the opencl handle itself so it could have released it while still being used in the sycl application so we needed to retain it but this PR removes the 1:1 mapping between UR handles and OpenCL handles so this should no longer be needed.
• Moves SetArgForLocalAccessor.cpp and InteropKernelEnqueue.cpp unittests to be DeprecatedFeatures/set_arg_interop.cpp e2e-test. For more context refer to this comment

UR PR

[kbenzie]

@intel/llvm-reviewers-runtime please review

[steffenlarsen]

It seems like this changes the behavior for all paths, i.e. also L0, HIP and CUDA. Is that safe?

[kbenzie]

I'll look into this.

[steffenlarsen]

According to the SYCL 2020 specification's OpenCL backend section:

The lifetime of a SYCL runtime class that encapsulates an OpenCL opaque type and the instance of that opaque type retrieved via the get_native() free function are not tied in either direction given correct usage of OpenCL reference counting. For example if a user were to retrieve a cl_command_queue instance from a SYCL queue instance and then immediately destroy the SYCL queue instance, the cl_command_queue instance is still valid. Or if a user were to construct a SYCL queue instance from a cl_command_queue instance and then immediately release the cl_command_queue instance, the SYCL queue instance is still valid.

Note that a SYCL runtime class that encapsulates an OpenCL opaque type is not responsible for any incorrect use of OpenCL reference counting outside of the SYCL runtime. For example if a user were to retrieve a cl_command_queue instance from a SYCL queue instance and then release the cl_command_queue instance more than once without any prior retain then the SYCL queue instance that the cl_command_queue instance was retrieved from is now undefined.

Based on that, it seems like the old behavior was correct. Unless the UR backend makes sure of this implicitly now?

[kbenzie]

I believe there were some lifetime issues exposed as part of these changes, I'll need to investigate this further since Omar has changed teams.