One inference YAML per model; other accels commented under masking

Add per-rate inference YAMLs (_4x/_6x/_8x); keep name.yaml as 4x alias

Sync inference YAMLs from projects (init2 target_acceleration from val-4x)

Fix init2 inference YAMLs: restore target_acceleration

Docs: install from GitHub via conda + editable .[dev], not PyPI

Add DIRECT JOSS citation to model card

Rewrite model card: clearer docs, dataset links, no checkpoint iters

Fix inference YAMLs: pin single R+ACS; document training schedule + how to override

Fix direct predict usage: output_directory is positional