feat: add dedicated Mistral AI provider with model generator

[Djiit]

Summary

Fixes #14

• Added Mistral AI provider with a generator script (cmd/mistral/main.go) that fetches models from the /v1/models API endpoint
• Uses openai-compat type (Mistral's API is OpenAI-compatible)
• Generates config with 11 curated models: Codestral, Devstral 2, Magistral Medium/Small (reasoning), Ministral 3B/8B/14B, Mistral Large 3, Medium 3.5, Small 4, and NeMo
• Added InferenceProviderMistral constant, provider registration, and gen:mistral Taskfile entry

Context

Supersedes #118. Addresses the reviewer feedback from #118 (comment) — uses a generator that fetches from Mistral's /v1/models API rather than a hand-crafted JSON.

Key design decisions:

• Allowlist approach: The Mistral API returns many aliases for the same model (e.g. mistral-small-2603, mistral-small-latest, magistral-small-latest are all the same model). The generator uses a curated wantedModels set of canonical -latest IDs to avoid duplicates.
• Hardcoded pricing: The /v1/models endpoint doesn't expose pricing, so costs are maintained in a map (sourced from https://mistral.ai/pricing).
• Reasoning support: Magistral models are marked with can_reason: true and reasoning levels.

Test plan

• go build ./... passes
• go vet ./... passes
• go test ./... passes (37 tests across 22 packages)
• go run cmd/mistral/main.go generates valid config with 11 models
• Test with crush: go run . then CATWALK_URL=http://localhost:8080/ crush

Template

• I have read CONTRIBUTING.md.
• I have created a discussion that was approved by a maintainer (for new features):