AMIP simulation core

[JGrassi97]

Scope: long AMIP simulations for NeuralGCM with robust forcing management and checkpointing.

What changed:

• Added RUN_TYPE separation (hindcast / amip) in simulation logic.
• Implemented full AMIP forcing ingestion from dedicated forcing source in sim_neuralgcm.py.
• Added explicit validation for forcing path and run mode consistency.
• Enabled chunked inference with checkpoint resume support for long runs.
• Added per-chunk netCDF writing to reduce RAM pressure and improve crash resilience.

Operational impact:

• AMIP runs are now stable for long horizons.
• Restarts are safe after interruption (checkpoint + chunk outputs).
• Memory usage is bounded during inference and output writing.

Commit anchors: 93f5b79, c1b3add, 8a170c3, 49ad166

---

Release v0.2.0-rc2 — Forcing & download reliability

Scope: simplify forcing workflow and avoid unnecessary data transfers.

What changed:

• Switched AMIP forcing workflow to pre-built monthly NetCDF forcing file.
• Updated config pathing for AMIP forcing input.
• Added early-exit guards in ERA5/EERIE download scripts when output folder is already populated.

Operational impact:

• Faster reruns on existing experiments.
• Less wasted bandwidth/time on repeated downloads.
• Clearer and more deterministic forcing source for AMIP runs.

Commit anchors: 536c65f, f50f088

---

Release v0.2.0-rc3 — Documentation & support visibility

Scope: make capability matrix immediately readable for users.

What changed:

• Added Supported modes column in support matrix docs.
• Updated README release banner and compatibility table content.

Operational impact:

• Users can quickly identify which model supports amip, hindcast, or both.

Commit anchors: 2df4715

---

Release Recommendation

Proposed merge tag: v0.2.0

Why this is release-ready:

• Core AMIP functionality implemented end-to-end.
• Reliability hardening completed (checkpointing, chunked outputs, rerun safety).
• Documentation aligned with actual supported modes.