feat(providers): provider-aware retry with exponential backoff and rate-limit handling

[EngineerProjects]

Summary

Providers like Anthropic and OpenAI return 429 Too Many Requests with
a retry-after header. The current retry logic does not honour this header
and uses fixed delays. This leads to wasted retries and unnecessary failures
on long multi-agent runs.

Scope

• Read retry-after / x-ratelimit-reset-requests headers on 429 responses
• Exponential backoff with jitter for 5xx errors
• Per-provider retry config: max retries, base delay, max delay
• Expose config.RetryPolicy in ClientConfig

Proposed config

type RetryPolicy struct {
    MaxRetries    int           // default 3
    BaseDelay     time.Duration // default 1s
    MaxDelay      time.Duration // default 60s
    HonorRetryAfter bool        // default true
}

Acceptance criteria

• RetryPolicy struct in pkg/sdk
• retry-after header parsed and respected (sleep until reset time)
• Exponential backoff with ±10% jitter for 5xx
• Unit tests: 429 with header → correct sleep duration; 5xx → backoff sequence
• No retry on 4xx errors other than 429