Follow-up on std.Target GPU changes in #20960

[alexrp]

• As discussed on Discord today, I discovered that my testing methodology in std.Target: A bunch of C type size/alignment fixes #20960 was flawed for GPUs due to using -xc instead of -xcl or -xhip, etc. So I'm reverting most of the GPU-related changes, save for 23b5a6c which is correct as far as I can tell.
  • I'm still not confident that the values that I'm restoring here were correct. For example, sizeof(long double) = 16 for AMDHSA seems quite suspicious. But having spent a whole day digging through this GPU nonsense and ending up not much wiser than when I started, I'm deciding to just leave it alone until someone actually exercises these code paths in the real world.
• I'm including a fix for C type alignment for SPIR-V targets. We had this wrong for sure, and I didn't catch it in std.Target: A bunch of C type size/alignment fixes #20960 (again, due to the flawed methodology).
• This also includes a fix to LLVM target triple construction for GPU backends. Surprisingly enough, the GPU backends actually care about the vendor tag in the target triple.
  • https://llvm.org/docs/AMDGPUUsage.html#target-triples
  • https://llvm.org/docs/DirectXUsage.html#target-triples
  • https://llvm.org/docs/NVPTXUsage.html#triples
  • https://llvm.org/docs/SPIRVUsage.html#target-triples
• I'm not restoring spir/spir64 after all. I can't find much evidence that this saw notable real world use even before the introduction of SPIR-V in 2015. Can always add them back if demand materializes.

[alexrp]

cc @Snektron