[Feature]: Auto-retry on MCP tool rate limit instead of terminating #16993

Problem

When an MCP tool call hits a rate limit, opencode immediately terminates the current task instead of retrying automatically.

Error message seen:

You have exceeded the request rate limit. Please wait a short period before retrying your request or contact the support to increase your request rate limit.

This is especially disruptive because the agent aborts mid-task, requiring the user to manually re-trigger the entire workflow from scratch.

Expected Behavior

When an MCP tool call returns a rate limit error (e.g. HTTP 429 or a message matching the pattern above), opencode should:

1. Automatically retry the tool call with exponential backoff instead of terminating
2. Inform the user that a rate limit was hit and a retry is in progress (e.g. Rate limited by MCP tool, retrying in Xs...)
3. Continue the task seamlessly after a successful retry

Current Behavior

opencode terminates the current generation/task immediately when an MCP tool responds with a rate limit error.

Context

• This is similar to the existing retry logic for AI model provider rate limits (see More resilient retry policy for Amazon Bedrock rate limiting #452)
• The same retry-with-backoff approach should be applied consistently to MCP tool calls
• Users should not need to restart their entire task just because an external tool was temporarily rate limited

Suggested Implementation

Apply the existing retry/backoff logic (used for LLM provider rate limits) to MCP tool call errors as well — detecting rate limit signals in the tool response and retrying with appropriate delay before propagating the error to the agent.