[replicate] Add verified model metadata

[rekram1-node]

Summary

• add the direct Replicate provider and the three currently resolvable models from the repository's Replicate discovery inventory
• mark DeepSeek R1 as fixed reasoning with no configurable reasoning controls
• omit reasoning options for the two non-reasoning Llama 3 deployments
• exclude the stale Meta Llama 3.1 405B deployment, which now returns 404

Evidence

• Replicate model-specific llms.txt schemas for all three included deployments
• checked-in Cloudflare AI Gateway Replicate discovery snapshot for inventory and pricing
• Replicate model URL availability checks, including the unavailable 405B ID

Validation

• bun validate
• git diff --check
• bun test: 43 passed, 1 baseline failure in the open-weight metadata test; the same 24 unrelated Sarvam/NVIDIA/Cohere missing-weight records fail on untouched origin/dev