Retraining and neural network emulator

This release includes improvements in the training of the emulator by using a discrete sine transform IR-resummation for the KECLEFT predictions that the emulator is made with respect to. This change leads to an approximate 30% improvement in accuracy, with no known issues.

We also include a new neural network version of the emulator that obviates the need to include KECLEFT predictions as input to the model, and this makes the predictions much faster.

v1.0-arxiv

This is the version of the code as implemented when the aemulus nu paper posted to the arxiv.