Adjust thresholds for bfloat16 product tests that are currently failing.

For bfloat16, the default epsilon for areNotApprox ends up being relatively large, causing some tests to fail even though practically the sides being compared are significantly different.

Also, some of the VERIFY_IS_APPROX tests when comparing matrices that are multiplications of 3 matrices are failing for bfloat16, since coefficients can grow quite large, but we don't adjust the threshold. Doubling the threshold seems to allow the test to pass reliably.