[mlir][tensor][bufferize] Fix dealloc placement in scf.forall op

The terminator of this op is special: it does not just yield a value,
but bufferizes to a memcpy. This requires special treatment to make sure
that deallocs are placed after the memcpy. (By default, deallocs are
placed right before the terminator.)

Differential Revision: https://reviews.llvm.org/D148408