Cmake revisions; uniform nums; testing scripts - hopefully working with travis

[drallensmith]

Let's go from simplest to most complicated (or at least most uncertain):

• I've modified the cmake material (README.md, CMakeLists.txt, and .travis.yml) to use - optionally for manual compilations - RelwithDebInfo (optimized but with debug information - details vary depending on platform and compiler) for this and the rcsc library, and to use MinSizeRel (optimized but try to avoid large sizes; e.g., -Os) for the server & soccerwindow2 (given the stadium.patch file, I figured these were less likely to need speed). It tests out fine locally, at least.
• I have put in uniform nums as part of the low-level state space. The form that they are in wound up a bit strange in some respects - the verification code wants it to be within -1 and 1, unless -2 for invalid, so I wound up having it return a -1 for unknown and unum/100.0 for known.
• I have created some testing scripts, one simple (albeit requiring a hfo.py change so that it could know exactly how many different action & status types to test) and one starting up the server and testing out interactions (using the low-level state space). This turned out to be particularly a good idea for the second one of the above, since most of the example agents use the high-level and as a result I didn't initially spot the validation code problem above. (Yes, the validation code could be removed, as it is for the high-level state space, but that's more likely to break things that expect all features to be within the standard range.) Figuring out a way to maximize the chance of the testing agent "seeing" the npc agents turned out to be a bit difficult, although I am currently working on some changes using an action version of doPreprocess that may help.
• The part I am least sure about, due to being unable to locally test it out, is expanding the travis testing to include the scripts. Hopefully it'll work OK when this pull request goes in; if not, I'll back out either the second script or both scripts from being tested via travis.

-Allen

[drallensmith]

The second test script, which actually starts up the server, appears to not work with travis due to the virtual environment in use (I have read of other instances of this problem, even with localhost). But at least the first test script can be run automatically.

-Allen

[mhauskn]

Thanks for the PR, these changes look good! In the test script you can enable --fullstate argument to HFO to check that unums come through correctly when there is no partial observability.

Next item is to update the manual to reflect the change in the low-level feature space.

[drallensmith]

Good point re --fullstate (although it was still interesting to see exactly how little information the learner will have at times) and the manual - will try to work on that when I have a bit of time (job interview coming up).