In overconstrained planning (nullable=true), CH FIRST_FIT must first evaluate value null (not last), in case that step is terminated before all values are evaluated

Following a chat discussion with ge0ffrey on freenode #optaplanner-dev I open a Jira ticket as requested.

I observe a Hard constraint violated without a score trap.

I give to Optaplanner the following simple cloud balancing problem to solve:
"try to place 1000 processes consuming "1" cpu on 75 hosts with 2 cpus each".
The correct solution is to place 150 processes but Optaplanner places 151 processes so it breaks a hard constraint (A host has 3 processes on it).

After rules analysis, ge0ffrey confirms that it looks like a bug.

Please find attached a unit test that reproduces the problem (using Optaplanner 6.4)
It runs on Windows but you can easily adapt it for Linux.