Add ability to set a constant random seed for shuffling in Batcher

[BradLarson]

PR #418 replaced the core usage of TensorFlow.Dataset with the new Batcher mechanism. This is a step forward, but there was one small piece of old functionality missing in the upgrade: the ability to set a constant seed for shuffling.

This was largely of interest to the Benchmarks target, to minimize noise due to slight variations in random shuffling from run to run. It isn't a huge issue, and won't have much of an impact on the benchmarks as they are performed today, but it would be nice to add this functionality back at some point. This is a tracking issue to remind us to do that.