Harden TriAttention model calibration preflight

[aivrar]

Summary

• validate the matching Hugging Face model context, nested RoPE metadata, and runtime-version difference before downloading weights
• estimate official one-shot calibration memory and support explicit per-GPU cuda:N selection
• add a compatibility runner for nested RoPE and selected-device handling, with UI and documentation updates

Mythos-nano-heretic / IQ4_XS findings

• IQ4_XS describes the inference GGUF; calibration still loads the source checkpoint in BF16
• the model declares a 131,072-token context, so a 200,000-token one-shot calibration is rejected
• its nested RoPE theta is 1,000,000, conflicting with the legacy 10,000 fallback exposed by the pinned runtime; the wrapper now selects the nested value
• a 200,000-token official run has an estimated 36.3 GiB memory floor and cannot fit a 24 GiB GPU

Verification

• python -m ruff check (changed Python files)
• python -m compileall -q multi_turboquant run_ui.py
• python -m pytest -q (443 passed, 14 skipped)