Some NFC changes to how we scan of kexts & kernels in memory in the

DynamicLoaderDarwinKernel plugin.  Created a new function ReadMachHeader
and instead of reading through the target cached memory reader,
start by reading only a mach header sized chunk of memory, then
check it for a valid mach-o magic # and use the size of the load
commands to pre-fetch the entire load commands of the kext which
is the only thing we're going to read, instead of letting the generic
mach-o parser read it in 512 byte chunks.

Functionally this is doing exactly the same thing as before, but by
cutting down on the # of packets going back and forth, even on a 
local connection it's close to a quarter faster than it was before.

<rdar://problem/38570146> 

llvm-svn: 334995