UR_PROGRAM_INFO_KERNEL_NAMES query is going to be difficult for some adapters to implement

There isn't a way for the CUDA or HIP adapters to use their backend APIs to query out the names of the entry points in a program object. There was a thought encoded in comments in the original PI implementation that the IR could be searched through for these (see here). This works if the IR is a text format like ptx is, but if it isn't or the program was created from a device binary that solution looks less likely to be worthwhile.

It's unclear how useful this property actually is to people, the above comment suggests SYCL is using it, but we obviously aren't hitting that sycl::detail::pi::die("getKernelNames not implemented"); so they must be working around this somehow. At the very least the kernel handling in the CTS will need redesigned without the assumption that it can be used, and we should probably make it optional in the ProgramGetInfo spec/cts test unless these adapters do come up with a nice way to implement it.

Start with the linked comments in the original PI implementation and inspect the ProgramGetInfo specification and its CTS test. Determine whether kernel-name queries can be supported across CUDA, HIP, text IR, and device-binary programs. Done means the CTS no longer assumes this property is universally available and the specification and test consistently reflect the decision.