Tighten floating-point comparison tolerances in tests.

[deansher]

As of this writing, we have tests with wide floating-point comparison tolerances. See, for example, call paths into TestSession.setEpsilon(...). @JimClarke5 reports, "I have seen a wide difference in the number of significant digits in other test cases when I use Java to determine the expected result."

[Craigacp]

Let's be careful with this though. I observe differences of 2% accuracy on mnist across different versions of Java with and without strict math in Tribuo, and math.exp gives different answers depending on java 8, java 9+ and running on x86_64 and arm64 platforms.