feat: update STACKIT models catalogue

[fheuer]

Summary

This PR refreshes the full STACKIT shared-models catalogue against the official docs Available Shared Models.
Prices adjusted to a rate of 1.17 (EUR→USD).

New models

• openai/gpt-oss-20b:
  • 131K context, 8192 max output
  • tool calling and reasoning enabled
  • $0.18/$0.29 per 1M
• Qwen/Qwen3.6-27B
  • linked via base_model = "alibaba/qwen3.6-27b" with STACKIT overrides
  • 262K context, 16384 max output
  • text-only input, reasoning and structured_output disabled
  • $0.53/$0.76 per 1M

Catalogue updates

• Repriced all live models using the 6-month average conversion rate of 1.17 provided by the (European Central Bank, ECB)
• Corrected max-output limits for these models:
  • Qwen3-VL 235B → 16384; Llama-3.3-70B and Gemma 3 27B → 4096
• Marked Mistral-Nemo and Llama 3.1 8B as status = "deprecated" (docs flag them
  "Deprecated - migrate to gpt-oss-20b")

Test plan

• Specs and prices verified against STACKIT docs (per-model facts tables)
• bun run validate passed
• base_model resolution confirmed (canonical inheritance, overrides applied)

[rekram1-node]

AUTOMATED REVIEW: Blocking merge: keep the documented 218K context for Qwen3-VL; delete the two NeuralMagic models that are no longer in the supported catalog rather than marking them deprecated; and add explicit reasoning_options = [] to the reasoning GPT-OSS 120B deployment.

[fheuer]

Hey Aiden, the automated review feedback is addressed:

• Context for Qwen3-VL adjusted to 218K
• Removed the previously marked as deprecated and now no longer supported models
• Specified reasoning_options for gpt-oss-120b
• Checked reasoning-enabled models for missing reasoning_options in TOML configs

@rekram1-node