Bugfix for #2017 broken across_channels computation in MVNLayer

[cdoersch]

The current implementation of MVNLayer requires multiplying the data matrix by a vector of 1's to perform sums for the mean and variance. This vector, sum_multiplier, gets created with the incorrect size when across_channels is true (it needs to be channels*width*height). See #2017. This PR checks for the across_channels flag and creates the sum_multiplier with the correct size.

Note that this is fixed in #1979 but it's looking less and less likely that this PR will be merged.

@seanbell can you confirm that this fixes all the bugs you're aware of in mvnlayer?

[longjon]

Thanks for breaking out this fix, @cdoersch, and to answer @jyegerlehner and @seanbell from #1979, that's the right thing to do in this situation.

@seanbell claimed this was not failing a test because the test was also incorrect; is that still true? Ideally this PR (and most(?) bug fix PRs) should have two commits: one that fixes the test and fails Travis, and one that corrects the bug and passes. That way we have good confidence that the fix won't regress in the future!

[jyegerlehner]

@cdoersch my recollection is there are a couple more. 1) the backward gradient calculation was wrong in the mean-only case. 2) The tests weren't testing what they purport to, because they were using Layer::ParseFromString instead of google::protobuf::TextFormat::ParseFromString to parse the mvn_param, so the prototxt was ignored. 3) Something else? I don't remember at the moment.

I created 2964 where I first fixed only the tests per @longjon suggestion, which should show about half the tests failing. I'll cherry-pick the MVNLayer fixes from 1979.

[cdoersch]

Looks like #2964 incorporates all these changes. Closing.