feat(proxy): make connection-level retries configurable per provider

[kianwoon]

Summary

• Add connectionRetries per-provider config (default 3) to control retry count before fallback escalation
• Reduces worst-case latency from ~35s to ~8-17s for operators who set lower values
• Backward compatible — omitting the field preserves current behavior (3 retries)

Files changed

• src/types.ts — _connectionRetries?: number on ProviderConfig
• src/config.ts — Zod schema, peekConfig type, loadConfig mapping
• src/proxy.ts — forwardWithRetry() reads provider._connectionRetries ?? 3

Config example

providers:
  glm:
    connectionRetries: 1   # fail fast, move to fallback quickly
  minimax:
    connectionRetries: 2   # slightly more tolerant

Motivation

From 2hr production log analysis: 108 connection errors on GLM with 3 retries each wasted ~35s per failed request before fallback. With connectionRetries: 1, that drops to ~8.5s.

Test plan

• Verify npm run build passes with no type errors
• Set connectionRetries: 0 on a provider and confirm it skips retries, goes straight to fallback
• Set connectionRetries: 1 and confirm exactly 1 retry before fallback
• Omit connectionRetries and confirm default behavior (3 retries) preserved