adk: nil event from custom agent is converted into panic error

**Describe the bug**

When a custom Agent yields a nil event through its AsyncIterator, the ADK flow wrapper dereferences the event while assigning RunPath. The nil pointer panic is recovered and exposed as a generic panic error with a stack trace.

This affects both the Message and AgenticMessage flow paths. For an Agent wrapped as an AgentTool, the caller receives the panic error instead of a clear contract error.

**To Reproduce**

1. Implement a custom Agent whose Run method sends nil before closing its AsyncGenerator.
2. Wrap it with NewAgentTool or execute it with Runner.
3. Consume the returned iterator.
4. Observe a recovered nil-pointer panic originating in adk/flow.go.

Minimal behavior:

```go
iter, gen := adk.NewAsyncIteratorPair[*adk.AgentEvent]()
go func() {
defer gen.Close()
gen.Send(nil)
}()
return iter
```

**Expected behavior**

The framework should reject a nil event at the central flow boundary and return an actionable error such as:

`agent 'name' returned nil event`

It should not rely on panic recovery for invalid custom-agent output.

**Version**

Current main before the fix: `9d983b36`

**Environment**

```
go version go1.25.12 darwin/arm64
GOOS=darwin
GOARCH=arm64
```

**Additional context**

I have a small fix with regression coverage for both Message and AgenticMessage paths. The full `go test ./adk/...` suite and focused race-detector tests pass.

Start in adk/flow.go at the flow boundary where RunPath is assigned, then trace the Message and AgenticMessage paths used by Runner and NewAgentTool. Add regression coverage for a custom agent sending a nil event, and run the focused race-detector tests plus go test ./adk/...; done means an actionable agent-name error is returned instead of a recovered panic.

Written by the indexing model from the issue text.