fix(anima): correct RoPE configuration to match the reference model (#9189)

Anima's transformer was configured with incorrect positional-embedding (RoPE) values that did not match the configuration the model was trained with. This reduced quality on every Anima generation (weaker structure and more artifacts).

Corrected to match the official Anima integration in diffusers (huggingface/diffusers#13732), specifically the Cosmos-2.0-Diffusion-2B-Text2Image configuration in scripts/convert_cosmos_to_diffusers.py (rope_scale=(1.0, 4.0, 4.0)). Anima reuses the Cosmos-Predict2 DiT, so that is the authoritative source for its transformer configuration.

Note: this changes Anima output; existing seeds/workflows will not reproduce prior (degraded) images.

Co-authored-by: Your Name <you@example.com>
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>