CCAdL's reproducibility was broken in branch WalkerEnsemble

[FrederikHeber]

CCAdL reproducibility regression test was working in branch Averages_burnin and was broken during the introduction of the ensemble of walkers in the huge branch WalkerEnsemble.

See the description in the regression test script. The probable cause is a deviation in the gamma, as it is adaptively tuned in each step. This would cause all values (kinetic energy, total energy, gradient, noise, ...) to change when gamma is slightly different.
Check whether total_noise which controls changes to gamma lacks a tf.control_dependency or some other cause ... it is strange though that the behavior is consistent for CCAdL' sampler test and changes with every run in the reproducibility test.

[FrederikHeber]

Note that the CCAdL regression test is also broken in all recent tensorflow versions except for 1.4.1 for which the test was generated. We can probably learn something by looking at the (different) output of the various versions.