Add autoencoder configs and update defaults

[sgreenbury]

Summary

Adds four dataset-specific autoencoder experiment configs under local_hydra/ and tunes three defaults in src/autocast/configs/ that are currently the right choices for the runs we're doing across datasets.

New local experiment configs (local_hydra/local_experiment/ae/<dataset>/ae_dc_large.yaml)

• advection_diffusion
• conditioned_navier_stokes
• gpe_laser_wake_only
• gray_scott

Default changes in src/autocast/configs/ (documented inline)

Config	Change	Restore previous behaviour
autoencoder.yaml	float32_matmul_precision: null -> high	float32_matmul_precision=null
datamodule/gpe_laser_only_wake.yaml	add channel_idxs: [1, 2] (real + imag of wave function)	~datamodule.channel_idxs
trainer/default.yaml	add best-val-{epoch}-{val_loss} ModelCheckpoint (top-1, monitors val_loss)	~trainer.callbacks.1

Each change has a short inline comment dated 2026-04-17 noting what changed and how to opt out. A TODO in trainer/default.yaml flags a follow-up to refactor callbacks into composable config groups so the best-val opt-out does not rely on a list index.

Note on dataset channel choices

The new channel_idxs: [1, 2] on gpe_laser_only_wake keeps the real and imaginary parts of the wavefunction (full state that the GPE simulator evolves).

Test plan

• CI passes
• Spot-check that the four new ae_dc_large.yaml configs resolve via uv run autocast ae --dry-run local_experiment=ae/<dataset>/ae_dc_large