Pin h5py version to 2.10.0.

[cmnbroad]

It looks like the conda env recently started resolving h5py to v3.1.0, which in turn appears to be incompatible with the keras version we're using, causing the CNNScoreVariants integration tests to fail when keras tries to load the model file (see tensorflow/tensorflow#44467). This PR pins the version to the version used by the last build I could find that succeeded, which is 2.10.0.

[cmnbroad]

Good point on the package version test. I'll pin r-backport and add package tests for those.

Perhaps we can forgo pinning everything, but add a check that fails if the conda environment has been changed in any way?

Some kind of alert when things change would be good, but we'd still need to pin in a situation like this I think?

[samuelklee]

Yes, definitely pin h5py in this case.

Not 100% sure what form the alert should take or what manual interventions should be required when it’s triggered, but perhaps we can discuss at the next dependency meeting.

[gatk-bot]

Travis reported job failures from build 32095
Failures in the following jobs:

Test Type	JDK	Job ID	Logs
conda	openjdk8	32095.5	logs

[droazen]

@cmnbroad @samuelklee Seems like there's at least one test failure on this branch:

PythonEnvironmentIntegrationTest. testGATKPythonEnvironmentPackagePresent
java.lang.AssertionError: The installed version of r-backports does not match the 1.1.10 version that was requested. Check the build log to see the actual version that was resolved by conda.

[samuelklee]

I would guess that we actually do get the requested r-backports 1.1.10, but since this is an R package rather than a python package, the python package version check performed by the failing test is not applicable (i.e., you cannot import r-backports in python).

[droazen]

@samuelklee That certainly makes sense -- I'll push a change to this branch to remove r-backports from the DataProvider in question