[Feature] Emit SSE event before/after compress_context() for better UX

## Problem

When `compress_context()` is triggered (token count exceeds threshold), the operation blocks silently with no SSE event emitted to the frontend. Users see a "thinking" spinner but don't know the system is actually compressing context.

Worse, if compression fails (e.g., `RuntimeError: Failed to generate structured output for model` due to unsupported `DataBlock` in context), the error is classified as `ErrorType.UNKNOWN` and surfaces as a generic "unknown error" to users.

## Reproduction

1. Have a long conversation that accumulates >52k tokens
2. Send a message that triggers `compress_context()`
3. Observe: no SSE event during compression, just silence
4. If compression fails: `REPLY_END(finished_reason=error, error.type=unknown)`

## Current Code

In `_agent.py:825-831`:
```python
if action == "reasoning":
# Compressed the memory if needed before reasoning
await self.compress_context() # ← blocking, no event
async for evt in self._inject_runtime_state():
yield evt
```

The `on_compress_context` middleware hook exists but has signature `-> None`, not `AsyncGenerator`, so it cannot yield events.

## Suggested Solution

Emit a `HintBlock` event before compression starts:

```python
if action == "reasoning":
if self._should_compress():
yield Msg(
name="system",
content=[HintBlock(hint="Compressing conversation context...")],
role="user",
)
await self.compress_context()
async for evt in self._inject_runtime_state():
yield evt
```

This would allow frontends to display "Compressing context..." during the operation.

## Environment

- AgentScope version: 2.0.4 (editable install from main branch)
- Model: DeepSeek V4
- Runtime: BusinessEmbedDemo

## Additional Context

The compression failure itself is a separate bug - `generate_structured_output()` doesn't handle `DataBlock` gracefully. But even if that's fixed, the UX issue of silent compression remains.