[Metal] Crash in AIRBuilder::patchClzCtz when compiling HLSL firstbitlow on 32-element vectors

We are seeing a failure on Metal via DXC but not clang-dxc likely because we scalarize clang-dxc.

https://feedbackassistant.apple.com/feedback/24450160

OffloadTest-mtl :: Feature/HLSLLib/long-vector/firstbitlow.long-vector.32.test

The issue is in libmetalirconverter.dylib. I am not on the newest shader converter I am on metal-irconverter version: 3.0.0

While vectorized DXIL is a new feature the metal-irconverter has been able to convert other vectorized instructions just fine. It is unclear what is unique in this case as it is an existing instruction just its vector form.

# | Stack dump without symbol names (ensure you have llvm-symbolizer in your PATH or set the environment var `LLVM_SYMBOLIZER_PATH` to point to it):
# | 0  offloader                 0x000000010259d208 llvm::sys::PrintStackTrace(llvm::raw_ostream&, int) + 56
# | 1  offloader                 0x000000010259af50 llvm::sys::RunSignalHandlers() + 204
# | 2  offloader                 0x000000010259dd08 SignalHandler(int, __siginfo*, void*) + 320
# | 3  libsystem_platform.dylib  0x000000018a843744 _sigtramp + 56
# | 4  libmetalirconverter.dylib 0x0000000106a2e774 AIRBuilder::patchClzCtz(llvm::MDNode*, llvm::Instruction&, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>&) + 308
# | 5  libmetalirconverter.dylib 0x0000000106a2e774 AIRBuilder::patchClzCtz(llvm::MDNode*, llvm::Instruction&, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>&) + 308
# | 6  offloader                 0x00000001025d7c7c (anonymous namespace)::MTLDevice::createPipelineCs(llvm::StringRef, offloadtest::BindingsDesc const&, offloadtest::ShaderContainer) + 212
# | 7  offloader                 0x00000001025d72c4 (anonymous namespace)::MTLDevice::executeProgram(offloadtest::Pipeline&) + 3868
# | 8  offloader                 0x0000000102540998 main + 2072
# | 9  dyld                      0x000000018a471d54 start + 7184
# `-----------------------------

This is the HLSL causing the crash:

StructuredBuffer<uint> In0 : register(t0);
RWStructuredBuffer<uint> Out : register(u1);

[numthreads(1,1,1)]
void main() {
  vector<uint, 5> Input0_5;
  for (uint I = 0; I < 5; ++I)
    Input0_5[I] = In0[0 + I];
  vector<uint, 5> Result5 = firstbitlow(Input0_5);
  for (uint I = 0; I < 5; ++I)
    Out[0 + I] = Result5[I];
}

Start with OffloadTest-mtl :: Feature/HLSLLib/long-vector/firstbitlow.long-vector.32.test and reproduce the crash using the supplied HLSL on Metal IR Converter 3.0.0. Compare the vectorized DXC path with clang-dxc, then determine whether the failure is actionable in this suite or belongs to libmetalirconverter; done means the named test no longer crashes.