fix a venom performance bug, where the stack would be reordered before
all jump instructions.

for "joining" jump instructions (where the target basic block can have
multiple cfg inputs), the stack reorder is needed, since we need the
invariant that the stack layout needs to be the same no matter which
basic block we jump to the target basic block from.

however, before jumping into a block with only a single `cfg_in`
(which, after cfg normalization, is equivalent to "splitting" jump
instructions `jnz` and `djmp`), the stack reorder is unneeded, since
stack reordering happens anyways in the target basic block, and no
invariant on the incoming stack layout is required.

this commit changes the behavior so that stack reordering only occurs
before these "joining" jump instructions.

on branch-heavy code, this can improve codesize by as much as 25%, with
corresponding gas improvement, especially in the presence of loops.