Add MEDS-EIC-AR as a registered model

Register MEDS-EIC-AR (MEDS-EIC-AR on PyPI) as a model in src/MEDS_DEV/models/. It's a MEDS-native autoregressive generative model with zero-shot inference capability — fits the benchmark's dataset-agnostic invariant since it takes MEDS data directly without per-dataset conversion rules (unlike #248 / cehrgpt, where the conversion-type flag was blocking).

What this needs

• src/MEDS_DEV/models/meds_eic_ar/ (probably the canonical name; PyPI dashes → underscores) with:
  • model.yaml declaring unsupervised: train (pretrain) and supervised: {train, predict} if the API supports fine-tuning + prediction. From the README, zero-shot inference is the headline use case — worth confirming whether MEDS-DEV's supervised lane fits or whether this is an unsupervised-only registration with a separate prediction path.
  • requirements.txt pinning MEDS-EIC-AR==<version> (latest stable).
  • README.md with a brief description + link to upstream.
  • refs.bib (Zenodo DOI is available: 10.5281/zenodo.17535558).

Open questions for the contributor

• Capacity variants — should we expose tiny / standard / large from day one (matching the meds_tab/tiny pattern), or land a single default and add sizes later?
• Optional dependencies (wandb, mlflow, flash-attn) — keep base install lean, mention extras in the README.
• Zero-shot use: does the MEDS-DEV supervised.predict flow apply, or do we need a new lane?

Pinging @mmcdermott (model author) for guidance on the right structure.

Refs

• Upstream: https://github.com/mmcdermott/MEDS_EIC_AR
• Existing model.yaml patterns for reference: cehrbert/model.yaml, meds_tab/tiny/model.yaml.

Start by reading the existing model.yaml patterns in cehrbert and meds_tab/tiny, then inspect the MEDS-EIC-AR README and PyPI release to confirm its API, latest stable version, and supervised or zero-shot support. Add the model directory with model.yaml, requirements.txt, README.md, and refs.bib, resolving capacity and optional-dependency choices; done means the registration matches a supported MEDS-DEV lane and references the upstream model.