I assumed that the set would vary (although using eval to convert the arrayref returned from parse into a closure like your tst could be interesting even if compile could end up longer than parse) and tried to optimize for the "many points" and "many ranges" cases; thus all those extra passes in the parse sub to simplify the result. I excluded floating point because exact comparisons are not reliable and it is easier to restrict the domain to integers. Multiple dimensions start to get complex enough that I suspect a general optimal solution is impossible — optimizing that case will depend on the actual data used.
The "suggesting but not providing an example" remark was an excuse for posting my solution — usually these questions have been "done to death" before I get to them. ;-)