fix(agentlang): isolate stream fallback correlations across agent contexts

[Linxiushen]

Problem

CorrelationIdManager already scopes paired event correlations by AgentContext.context_id, but the V2 stream-to-regular fallback correlation is still stored in a single process-wide _stream_fallback_cid slot.

This becomes unsafe when a parent agent and one or more background sub-agents make LLM calls concurrently:

1. The parent stream stores its request ID for a possible regular-call fallback.
2. A sub-agent stream overwrites that process-wide value with its own request ID.
3. If the parent falls back first, it consumes the sub-agent's correlation ID.
4. The sub-agent then either loses its fallback correlation or consumes a value from another run.

The V2 client uses this correlation ID to merge streamed chunks with the final non-streamed response. A cross-context mix-up can therefore associate a fallback response with the wrong Agent run and breaks trace integrity for concurrent delegation.

Proposed change

• Store fallback correlation IDs by the existing AgentContext scope instead of introducing a new identifier.
• Pass the same context_id scope when setting, clearing, and consuming fallback correlations.
• Preserve the current global-scope behavior for callers without an AgentContext.
• Add regression tests for interleaved parent/sub-agent set, clear, and pop operations.

Acceptance criteria

• Concurrent AgentContexts cannot overwrite or consume each other's stream fallback correlation IDs.
• Clearing a successful stream only clears the marker for that AgentContext.
• Existing callers that omit a scope retain the current single global-slot semantics.
• Focused Agentlang tests pass.