feat(plugin): progressive plugin loading — activate plugin components on demand

[Luhaozhu]

Summary

Installed plugins used to inject all of their bundled capabilities into every request — one skill-list line per skill in the system prompt, plus the full JSON schema of every MCP tool in the tools parameter, re-sent on every ReAct round. TTFT therefore grew linearly with the number of installed plugins.

This PR introduces progressive plugin loading: the plugin becomes a runtime aggregation unit again.

• Assembly time: plugin components are subtracted from the run's enabled sets. The system prompt carries only a one-line directory entry per plugin (slug + description) and a single load_plugin tool is registered.
• Activation time: when the model decides a request matches a plugin's description, it calls load_plugin. Only then are the plugin's MCP servers connected (stateless HTTP clients appended to the live toolkit) and its skills registered. Schemas and skill instructions are recomputed each round, so everything is usable from the next step.
• Stickiness: activation is persisted per chat (chat_sessions.metadata.activated_plugins), so subsequent turns assemble the plugin eagerly with no repeated activation. Explicitly invoking a plugin (slash / + menu) counts as activation.
• Sub-agents: plugins bound to a sub-agent are deferred the same way, with run-scoped activation only (no persistence, to avoid leaking into the parent chat's assembly).

The plugin directory is sorted by slug and byte-stable regardless of activation state, so provider-side prefix caching pays for an activation only on the activation turn and the one after it.

Boundaries

• Plugins pinned by a restricted chat mode stay eagerly loaded (deliberate admin narrowing).
• Plugins with stdio-transport MCP components are not deferred.
• PLUGIN_PROGRESSIVE_LOADING=false reverts to eager assembly.

Also included

• Hide AgentScope's builtin Skill viewer schema: skills load exclusively through view_text_file in this stack, so the schema was pure per-round prefill waste and a wrong entry point bypassing path mapping / runtime hints / observability.
• Chat UI polish for the load_plugin tool call: localized label, inline plugin slug, and a formatted activation-summary card instead of a raw JSON dump.

Testing

• tests/test_plugin_progressive_loading.py (9 cases): deferral resolution (visibility / enabled intersection / stdio exclusion / activation & invocation exemption with persistence), directory rendering stability, load_plugin runtime side effects (toolkit mutation, allow rules, close-list, chat persistence, sub-agent no-persist), env kill switch.
• Existing plugin / chat-mode / citation regression suites pass.