fixed bug in runCrossValidate for merging of MCMC configurations

[danielturek]

Fixes a bug in runCrossValidate, to correctly merge the sampler configuration lists from two different MCMC configurations.

Used conf$addSampler as was proposed by @perrydv, to do this in a robust manner.

Also added a test (using the dyes example), where cross validation now gives results that seem reasonable. The test checks for the exact numerical cross validation results.

[danielturek]

@paciorek I suspect this will initially fail testing, due to the travis bug. But once that's fixed, this PR should pass (as it works for the cross-validation tests), and I'm hoping you can merge it into the upcoming release.

[perrydv]

@paciorek Does this test failure look like just the travis issue we've been seeing?

[paciorek]

Yeah, that is it. It didn't finish the other tests after numericTypes, but presumably those won't be an issue.

I should be fixing the precision error this morning.

[paciorek]

@danielturek I'm trying to look into an issue raised on the users list about runCrossValidate and from what I can see in looking at that, it looks like the bug fix for runCrossValidate in this PR introduced a bug where the MCMC for all but the predictive loss is being done ONLY over the predictive nodes and not the parameters. Note that additional predictive samplers are assigned to the original conf configuration, but then the MCMC is built from modelMCMCConf, which contains only the predictive samplers.

[danielturek]

@paciorek It appears this bug (embarrassing mistake) is mixing the changes made to conf versus modelMCMCConf. Is that your assessment also?

I'll open a PR fixing this as I believe it should be, but will wait for confirmation.

[paciorek]

Yes, that's what I see.

[danielturek]

This is being addressed in PR #1298.

Pending testing results.