RFR(M): arraycopy/clone/copyOf intrinsics fixes for Lworld

[Roland Westrelin]

http://cr.openjdk.java.net/~roland/valhalla/arraycopy-lworld/webrev.00/

The biggest part of the change is new tests and checks to go to the slow
path if one of source or destination is flattened but the other one is
not. There are also a few fixes not directly related to the
intrinsics. The trickiest part is that when executing copyOf for an
array of unknown type, we go to the slow path if the array is a non
flattened array of value types so that it's properly initialized with
default values. The alloc site for such an array is marked ar parse time
with a new flag.

Roland.