Drop manual bf16 handling (currently just in LLVMGPU) (#20313)

Get rid of the manual bfloat expansion pass because LLVM has a bfloat16
type and the AMDGU backend supports it.

The pass remains in the LLVMCPU pipeline (since not having it causes
linker errors that aren't worth fixing) and in the SPIR-V backend, where
there's no official bf16 type.

This PR didn't appear to cause compilation failures on CUDA

Author: krzysz00

compiler/src/iree/compiler/Codegen/LLVMGPU/Passes.cpp

@@ -1107,9 +1107,6 @@ static void addLowerToLLVMGPUPasses(OpPassManager &modulePassManager,
       .addPass(createCSEPass)
       // Handle complex operation conversion.
       .addPass(createConvertComplexToStandardPass)
-      // Convert BF16 operations to occur as F32.
-      .addPass(createConvertBf16ArithToF32Pass)
-      .addPass(createConvertBf16ToUInt16BuffersPass)
       // Math dialect ops rewrites, approximations, casts.
       .addPass(createMathTransformPass)
       .addPass(memref::createExpandOpsPass)