Various improvements including:
- Adding the 1600 dim cox problem
- Improving general evaluation
- Adding gmm_v1
- Adding better eval for Many Well problem
- Adding flexible
point_is_valid_fn - Bounding real_nvp scaling (greatly improves training stability)
- Switching from haiku to flax