Fix top of LLVM, and remove upper limit of LLVM version from CMakeLists.

[mcourteaux]

From CMake documentation:

The [version] argument requests a version with which the package found should be compatible. There are two possible forms in which it may be specified:

• A single version with the format major[.minor[.patch[.tweak]]], where each component is a numeric value.
• A version range with the format versionMin...[<]versionMax where versionMin and versionMax have the same format and constraints on components being integers as the single version. By default, both end points are included. By specifying <, the upper end point will be excluded. Version ranges are only supported with CMake 3.19 or later.

[mcourteaux]

For the failing builders, I don't know. I'm at a loss. CUDA not found?

[mcourteaux]

Regarding the CUDA errors:

Error: CUDA error: CUDA_ERROR_NOT_FOUND cuModuleGetFunction failed

Is something reported from halide_cuda_run(). I suspect that this means that the kernel function name mangling changed due to some changes in LLVM? We'd be able to debug this using HL_DEBUG_CODEGEN, but I don't wanna compile LLVM git on my machine for this test. Can someone ssh into that machine and run one test with HL_DEBUG_CODEGEN=1? @alexreinking @abadams

[abadams]

I don't believe the failures are related and we can't test fixes for them on main without the changes in this PR, so I'm going to merge this then work on the failures.