simplify op_dump() / -Dx sequencing
Currently, whenever we dump an op tree, we first call sequence(),
which walks the tree, creating address => sequence# mappings in
PL_op_sequence. Then when individual ops or op-next fields are displayed,
the sequence is looked up.
Instead, do away with the initial walk, and just map addresses on request.
This simplifies the code.
As a deliberate side-effect, it no longer assigns a seq# of zero to
null ops. This makes it easer to work out what's going on when you
call op_dump() during a debugging session with partially constructed
op-trees. It also removes the ambiguity in "====> 0" as to whether
op_next is NULL or just points to an op_null.