feat: improve proxy compatibility and fallback handling

[mxyhi]

Summary

• add same-provider upstream cooldown selection to avoid repeatedly hitting recently failed upstreams
• improve models routing compatibility for OpenAI, Gemini, and Anthropic request patterns
• close the Anthropic proxy loop by fixing auth fallback and stripping Anthropic-only query/header metadata on non-Anthropic fallbacks
• update OpenCode model labels and README quick-start notes

Details

• P0: add upstream_selector runtime and cooldown-aware ordering within the same provider/priority group
• P1: support explicit /v1/models*, /v1beta/models*, and /v1beta/openai/models* routing, plus header/query-aware provider hints
• P2: harden Anthropic compatibility for:
  • lowercase bearer local auth
  • Authorization -> upstream x-api-key fallback
  • stripping beta=true, anthropic-version, anthropic-beta, and x-stainless-* when Anthropic requests fall back to non-Anthropic providers
• docs/UI: keep .reference and planning files ignored, add OpenCode config guidance, and use readable OpenCode model display names

Validation

• cargo test -p token_proxy_core
• cargo test -p token_proxy --lib