Don't flatten test sets to keep track of agent id

[yahelnachum]

Resolution Table

Date	Branch	Link	Details
-	-	-	-
-	-	-	-

---

Problem

Right now we flatten out the test sets, the same as we do for the training sets. This means we lose the track_id of the agent. Instead we should not flatten out the tests sets so that we can predict all 50 future frames per each test agent.

Solution

Same as above.