HIP: force max threads per block to be 1024

[fxzjshm]

Some old compilers still use 256. Explicitly set it to 1024 to get correct result from ops like ARGMAX and GROUP_NORM.

Related: #10610, #11619

CC @IMbackK

[IMbackK]

over all i am fine with this, but will defer to @slaren on if this kind of vendor behavior is something we want to support (see discussion in #11619)

[IMbackK]

please move this down a bit as the find_package calls and the version check below are logically related operations

[slaren]

I saw the discussion, but don't have any knowledge about HIP/ROCm to have an opinion about this. If you think that it is not likely to cause issues to other users, feel free to merge it.

[fxzjshm]

@IMbackK Moved. Is this place proper?

@slaren This compiler flag is documented at https://clang.llvm.org/docs/ClangCommandLineReference.html#cmdoption-clang-gpu-max-threads-per-block. I've also compiled with ROCm 6.3.1 and no compile error is given, now testing test-backend-ops.

Update: test-backend-ops w/ ROCm 6.3.1 on gfx1100 passed.

[IMbackK]

This just resets the default, also maximum value for all current amd gpus, it dosent change the code generation at all on sane versions of llvm at this time. We might run in to problems in the future if amd changes this for a new gpu arch - but i think this is an acceptable risk.