Refactor: Decouple load balance metrics collection from DividePlugin

### Is there an existing issue for this?

- [x] I have searched the existing issues

### Current Behavior

Currently, DividePlugin contains strategy-specific callback logic for P2C and ShortestResponse load balance algorithms:
```java
if (ruleHandle.getLoadBalance().equals(P2C)) {
return chain.execute(exchange)
.doOnSuccess(e -> responseTrigger(upstream))
.doOnError(throwable -> responseTrigger(upstream));
} else if (ruleHandle.getLoadBalance().equals(SHORTEST_RESPONSE)) {
beginTime = System.currentTimeMillis();
return chain.execute(exchange)
.doOnSuccess(e -> successResponseTrigger(upstream));
}
return chain.execute(exchange);
```
This design has several issues:
1. Violates Single Responsibility Principle: DividePlugin's core responsibility is selecting an upstream and writing routing info into exchange. Metrics collection for specific load balance strategies should not be mixed in.
2. Violates Open/Closed Principle: Adding a new load balance strategy that requires runtime metrics (e.g., adaptive load balancing) would require modifying DividePlugin, rather than simply implementing a new LoadBalance class.
3. Thread safety issue: beginTime is an instance variable shared across concurrent requests, which is not thread-safe. It should be stored per-request (e.g., in exchange attributes).

### Expected Behavior

Add onSuccess and onError callback methods to the LoadBalance interface with default empty implementations:
```java
public interface LoadBalance {
Upstream select(List upstreamList, LoadBalanceData data);

default void onSuccess(Upstream upstream) {}

default void onError(Upstream upstream) {}
}

return chain.execute(exchange)
.doOnSuccess(e -> loadBalance.onSuccess(upstream))
.doOnError(t -> loadBalance.onError(upstream));
```

### Steps To Reproduce

_No response_

### Environment

```markdown
ShenYu version(s):latest
```

### Debug logs

_No response_

### Anything else?

i will do it

Start with DividePlugin and the LoadBalance interface, then trace the P2C and SHORTEST_RESPONSE callback paths described in the issue. Check how beginTime is stored across concurrent requests and review the affected LoadBalance implementations. Done means strategy-specific callbacks are decoupled from DividePlugin, request state is not shared, and relevant gateway tests pass.

Written by the indexing model from the issue text.