Changes in version 0.1.0                        

  - Implements the Limite decoder architecture directly with R torch.
  - Loads the published Base, Base Soup, and Violetto safetensors
    checkpoints and tokenizers from Hugging Face.
  - Provides mixed global and sliding-window grouped-query attention,
    partial RoPE, value embeddings, XSA, MUDD history mixing, learned
    residual coefficients, and tied sigmoid-softcapped logits.
  - Provides cached token generation and the official single-turn
    message format used by Violetto for mathematical reasoning.