Redesign evaluator

[luoyanjie]

No description provided.

[fonglh]

Key considerations

1. Performance. Possible due to poll waiting, each evaluation takes a minimum of 7 seconds or so.
   This should be reduced to 3 seconds, preferably faster.
2. Test independence. A crash (killed due to limits, bad code etc) on ANY of the test cases means no test report file, so all the test cases look like they've failed. This is misleading to both the student and the TAs.

1 possible solution is to discover all the test cases before hand, then run them 1 by 1. Overhead needs to be benchmarked.

[fonglh]

Polling was eliminated in #2308 and test types are now run independently (#2596, #2603, #2593, along with changes to the evaluator script in the images (Coursemology/evaluator-images#18))

Test case dependence is a necessary teaching aid, and running the test cases 1 by 1 is likely to lead to large overhead because of the cost of invoking and booting the test framework.

Solution would be to run by test group, then run independently only if the test group has a fatal crash. Show a warning to the user that the tests were run independently, in case the instructor relies on dependent tests.