feat(service): support sharing agent with MCPs and skills

**Background**

As a follow-up to #1844, while we already support sharing credentials, agents, and knowledge bases, the current agent sharing mechanism is incomplete.

At the moment, only the basic agent configuration is shared — i.e., `name`, `system prompt`, and related configs. However, a well-tuned agent typically also includes MCPs and skills (even memory), which are equally important and should be carried over during sharing. Without them, a shared agent loses much of its value.

**The Challenge**

MCPs, skills, and even memory files are all persisted within the agent's workspace(sandbox filesystem). This raises a fundamental question: should sharing an agent also involve sharing its workspace?

The current answer is **no**. The isolation model in our workspace manager is hierarchical: `user_id` is the top-level boundary, with `agent_id` and `session_id` nested within. This means when a user receives a shared agent and creates a new session, a **new workspace is created under their own account** — which is the correct behavior, as sharing a workspace across users would risk data interference.

However, this creates a gap: **the original agent's MCPs and skills need to be carried over into the new user's workspace**, and there is currently no mechanism to do so.

**Proposed Solution (Under Discussion)**

One candidate solution is to allow an agent to declare a set of **default MCPs and skills** as part of its configuration. When a brand-new workspace is created for that agent (regardless of isolation level), these defaults would be automatically applied as the initial setup.

However, this approach has known limitations:
1. MCP configs may contain sensitive information (e.g. API keys) and therefore cannot be naively copied across users
2. Skills present an additional challenge — we currently lack a dedicated file service for skills (though we do have blob abstraction and S3 integration for the RAG pipeline, it is not designed to serve skills)