For jruby#3173 and to check against future regression. Was broken in
9.0.0.0 but seems to be fixed in 9.0.1.0.

DCE used to do additional tests to decide if an instruction can be
deleted based on the scope's state. This is a no-no. DCE should
not do any additional tests about deletability based on scope
state (which was in reality liveness tests) -- it should simply use
whatever info is produced by LVA. Those tests should have been
part of LVA instead.

This really didn't affect correctness of DCE, but it would be
providing incorrect information about vars still live when the
scope was exited. Specifically, AddLocalVarLoadStoreInstructions pass
uses that LVA information to determine whether to add lvar stores
on exit. Without accurate LVA information, this can inadvertently
prevent lvar updates from being seen outside the scope.

This is part 1 of the fix for #3173.

The following fixes were made:

1. Same kind of fix as in d59eeec but applied to this pass.
   The actual addition of binding stores shouldn't do additional
   analysis not already present in the dataflow analysis. It should
   simply use information from that analysis and add stores, where
   required.

2. A specific outcome of 1. above is: the addition should not look
   at scope type while adding stores on exit. If the analysis says:
   "add a store", add a store. No need to do it for closures only.

3. There was a subtle bug in the code. For exit BBs, we were
   intersecting lvars that were dirtied with lvars that were live
   at that point. However, we were looking for 'varsLiveOnExit'
   which was erroneous. LVA is a backwards propagation problem
   and we should be looking at 'varsLiveOnEntry' (since the exit
   of the scope corresponds to the dataflow entry for LVA).

Between these fixes and the fix in 8c48c3bc, #3173 should be solved.