Add worst-case math library inputs to JAX numerics accuracy tests.

Add the single- and double-precision worst-case inputs compiled by Gladman et al. ("Accuracy of Mathematical Functions in Single, Double, Double Extended, and Quadruple Precision", https://members.loria.fr/PZimmermann/papers/accuracy.pdf) to interesting_points across the JAX math accuracy tests so sampled float32 and float64 runs always check known hard-to-round arguments.

Also raise five float64 error bounds by 0.5 ULP where these inputs expose larger underlying library errors:
- sinh on CPU: 1.5 -> 2.0 ULP
- acos on GPU: 1.5 -> 2.0 ULP
- acosh on GPU: 2.5 -> 3.0 ULP
- atan2 on GPU: 1.5 -> 2.0 ULP
- cos on GPU: 1.5 -> 2.0 ULP