Support for CUDA/HIP on OpenCL/Level0 via chipStar

[ColonelPhantom]

Very hacky support for chipStar, which is a compiler to run HIP/CUDA on OpenCL/Level Zero (currently fairly Intel focused).

Requires:

• oneAPI Basekit 2023.2 (2024.0 will not work)
• chipStar (https://github.com/CHIP-SPV/chipStar)
• H4I-HipBLAS (https://github.com/CHIP-SPV/H4I-HipBLAS)
• H4I-MklSHIM (https://github.com/CHIP-SPV/H4I-MKLShim)

Currently only supports building via Makefile, as I didn't yet look at CMake integration for these.

I also wrapped some stuff using unsupported hipBLAS functions in #ifdefs so they won't break the build, and sacrificed use_tensor_cores in the process.

Finally, I changed some local sizes from 1024 to 256 so it can run on my UHD 620 laptop. (I think modern Intel graphics support local sizes of up to 1024, so this isn't strictly necessary for those).

Performance is still much worse than CPU (i5-8350U), but seems a bit better than OpenCL from preliminary testing (not a proper benchmark run). Would love if someone with more powerful Intel hardware could give this a try!

[Jacoby1218]

Interested in how this runs on my ARC A770 16GB. There's already another SYCL project as well here: #2690

[ColonelPhantom]

Yeah, I'm aware of that PR. The advantage of chipStar over dpct is that it works with the same source file, instead of a separate one that has to remain in-sync with the original (unless you want to use SYCL on all platforms).

(Also, chipStar does not use SYCL, it uses the lower-level Level Zero or OpenCL, which are also used by SYCL. It's essentially a compiler from CUDA C++ to SPIR-V, whereas icpx is a compiler from SYCL to SPIR-V.)