feat: optimize auto routing

[eigmax]

feat: intelligent model fallback with cross-provider routing

When a model fails (context window exceeded, billing errors, API failures),
the router now automatically retries with alternative models instead of
showing raw API errors to users.

Key behaviors:

• Context window exceeded → escalate to higher-tier model with larger context
• Insufficient credits (402) → try same-tier cheaper alternatives first,
  then escalate to next tier
• Cross-provider fallback: if OpenRouter credits are exhausted, automatically
  route to direct Anthropic/Groq/OpenAI APIs using separately configured keys
• Pre-check: estimate token count before API call and skip models whose
  context window is too small
• User-friendly error messages when all fallbacks are exhausted

Changes:

• catalog.rs: parse context_length from OpenRouter API, single Catalog struct
• config.rs: add tier escalation chain (next_tier/prev_tier), tier_alternatives
  with models from multiple providers (Groq, DeepSeek, Anthropic, OpenAI)
• router.rs: expose get_context_length, get_fallback_model, get_tier_alternatives
• metrics.rs: track escalation_count
• litellm_provider.py: classify billing vs context errors, set all provider
  API keys, skip openrouter/ prefix for direct-provider calls
• loop.py: pre-check + escalation loop with billing-aware routing
• schema.py: add get_all_api_keys() for multi-provider support