We do not need a memory copying loop anymore in insn_fetch; we
can use a byte-aligned pointer to access instruction fields directly
from the fetch_cache.  This eliminates 50-150 cycles (corresponding to
a 5-10% improvement in performance) from each instruction.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>