[mlir] Don't make the ROCm conversions depend on the execution engine

During a conversion to MLIR_ENABLE_EXECUTION_ENGINE from checking for
the native target, the ROCm conversion passes (--serialize-to-hsaco)
were mistakenly flagged for being disabled if the execution ending is
not being built.

These passes use LLVM to build binaries for AMD GPUs, and so require
that backend to be enabled. However, they do not produce native code,
nor do they interact with the JIT or any of the execution engine
support libraries.

When building MLIR into a compiler library that's intended to produce
GPU binaries, we want to build only the AMDGPU backend and have the
binary serialization passes available. This change makes that
possible.

It looks like the CUDA path might currently require a native target,
it's hard to tell, so this commit leaves that if statement untouched.

Reviewed By: fmorac

Differential Revision: https://reviews.llvm.org/D155227