This change introduces umbrella headers for the files in the cpu subdirectories:

src/share/vm/utilities/bytes.hpp
src/share/vm/opto/ad.hpp
src/share/vm/code/nativeInst.hpp
src/share/vm/code/vmreg.inline.hpp
src/share/vm/interpreter/interp_masm.hpp

It also cleans up the include cascades for adGlobals*.hpp,
jniTypes*.hpp, vm_version*.hpp and register*.hpp.

Where possible, this change avoids includes in headers.
Eventually it adds a forward declaration.

vmreg_<cpu>.inline.hpp contains functions declared in register_cpu.hpp
and vmreg.hpp, so there is no obvious mapping to the shared files.
Still, I did not split the files in the cpu directories, as they are
rather small.

I didn't introduce a file for adGlobals_<cpu>.hpp, as adGlobals mainly
contains machine dependent, c2 specific register information. So I
think optoreg.hpp is a good header to place the adGlobals_<cpu>.hpp includes in,
and then use optoreg.hpp where symbols from adGlobals are needed.

I moved the constructor and destructor of CodeletMark to the .cpp
file, I don't think this is performance relevant. But having them in
the header requirs to pull interp_masm.hpp into interpreter.hpp, and
thus all the assembler include headers into a lot of files.