I thought I could get away without since I wasn't (presently) targeting
systems without hardware support for unaligned accesses, but on armv7
clang insists on optimizing into the one ARM instruction that requires
alignment anyway - LDM/STM.  Oops.  Damnit, clang.

Commit transform-dis-cases-x86_64.S, which I forgot to earlier.

The transformed code was incorrect because it assumed the pointer it was
writing to was where the code would execute, but it was actually
'rewritten_temp'.  Changed transform_dis_main to take a pc_trampoline
pointer, which also helps the test harness.  However, this means that it
has to be called after the trampoline has been allocated, while before
the trampoline allocation depended on the generated size; this change
doesn't bother to use two passes or anything, but just allocates a new
code buffer if the maximum possible size isn't available - not the end
of the world, since trampoline_ptr will still only be increased by the
actual size before the next hook in the series (if any).