Add subtarget caches to aarch64, arm, ppc, and x86.

These will make it easier to test further changes to the
code generation and optimization pipelines as those are
moved to subtargets initialized with target feature and
target cpu.

llvm-svn: 219106