Add get_safe_autocast_context helper that only enters torch.autocast on
devices supporting AMP (cuda, xpu, cpu) and falls back to a no-op on mps
and unknown backends. Route the previously unconditional/underspecified
autocasts (vla_jepa, groot, molmoact2, lerobot_eval) through it so
autocast can be requested unconditionally without breaking on unsupported
devices.
CarolinePascal