Add built-in structured retry + rate-limit handling for onchain actions (RPC + CDP)

### Language Implementation

- [ ] Python
- [ ] TypeScript

### Feature Type

- [ ] Action Provider Template
- [ ] Wallet Provider Template
- [ ] Framework Extension
- [ ] Core Requirements
- [x] Other

### 🚀 The feature, motivation and pitch

## Overview
AgentKit agents often perform onchain actions (reads, writes, balance checks, swaps) through RPC endpoints and/or CDP-backed services. In real deployments, these calls frequently hit transient failures (timeouts, 5xx), rate limits (429), or provider-specific throttling. Today, each agent implementation typically re-invents retry logic, backoff, and error categorization.

## Expected Behavior
- Transient errors are retried automatically with safe defaults.
- 429 responses respect `Retry-After` (when provided) and apply backoff.
- Non-retryable errors fail fast with clear error types.
- Retry behavior is configurable globally and per action provider.
- Developers can opt out completely (e.g., `retries: 0`) if desired.

## Steps to reproduce
1. Build an agent that repeatedly queries balances and submits a transaction.
2. Run it against an RPC endpoint with throttling (or under heavy load).
3. Observe intermittent timeouts / 429s cause the agent to fail and stop the task.
4. Add custom retries in userland and notice duplication across actions.

### Alternatives

## Proposed Solution
- Introduce a shared resilience layer used by all AgentKit network-bound actions (RPC, CDP, and future providers).
- Implement configurable retry policies:
- retry count
- backoff strategy (exponential with jitter)
- maximum delay
- optional respect for `Retry-After` headers
- Classify errors into retryable vs non-retryable (e.g., timeouts/5xx/429 vs validation/config errors).
- Expose typed errors such as `RateLimitedError`, `TransientNetworkError`, and `NonRetryableError`.
- Allow configuration at two levels:
- global defaults during `AgentKit` initialization
- per-provider overrides where needed
- Add optional instrumentation hooks (e.g., `onRetry`) for logging and telemetry.

## Additional Information
- This change reduces duplicated retry logic across agent implementations.
- Improves reliability for long-running or autonomous agents in production.
- Aligns AgentKit behavior with best practices for resilient API and RPC clients.
- Provides a foundation for future observability and metrics integrations.