Keep load_skill tool bound; remove dynamic tool gating and refine active-skill handling

[weilixu]

Motivation

• Avoid mutating self.graph.tools at runtime and keep the load_skill tool bound to the agent at all times to simplify behavior and avoid errors when self.graph lacks a tools attribute.
• Ensure active skill text is injected into LLM inputs as an internal system message and that skill outputs are captured correctly during both invoke and streaming flows.

Description

• Removed dynamic tool gating: deleted _apply_active_skill_tool_gating() and any code paths that mutated self.graph.tools, so load_skill remains statically available when skills are enabled.
• Added automa_ai/skills/active_skill.py with ActiveSkillState, ACTIVE_SKILL_HEADER, and format_active_skill_message() to manage active-skill state and formatting.
• Integrated active-skill state into the agent by adding _active_skill_state, injecting formatted active-skill system messages in _build_stream_inputs() ahead of retriever/memory context, and adding _capture_skill_tool_outputs() to mark skills loaded from tool responses.
• Added clear_active_skill() and request_clear_active_skill() helpers and adjusted invoke/stream flows to build inputs using the active-skill injection and to capture load_skill tool outputs from both synchronous responses and streaming ToolMessages.

Testing

• Ran pytest automa_ai/agents/langgraph_chatagent_test.py in this environment and test collection failed with ModuleNotFoundError: No module named 'langchain_core', so automated tests could not be executed here.
• Updated unit tests in automa_ai/agents/langgraph_chatagent_test.py to remove gating-related assertions and to verify active-skill injection, ordering, absence when unset, and capture of load_skill outputs, which should be exercised in a local environment where langchain_core is available.

---

Codex Task