[ML] Improve robustness to very low variance data

[tveasey]

This fixes a couple of potential issues:

1. If the variance of the data was zero it could cause a nan in the decomposition component errors. The most serious side effect is that it would cause us to fail to restore state on Windows.
2. The check for low coefficient of variation data in the periodic component test should always use the raw bucket values to calculate the mean.

This turned up in a unit test failure back porting #198 to 6.5. As such this is effectively a forward port of the fix I made directly to that commit. I've marked this as a non-issue as this is related to unreleased code.

[edsavage]

LGTM