Add a RAG MCP server for KubePlus documentation, for use with kagent agents

### Summary
Agents built on kagent currently connect to KubePlus via a basic MCP server exposing
imperative tools (kubectl-metrics, etc.). This issue proposes adding a second,
retrieval-focused MCP server that lets an agent ground its answers in KubePlus's own
docs and example manifests, rather than relying on model memory alone.

### Motivation
- kagent supports attaching multiple `ToolServer` resources to a single `Agent`, so
tools and retrieval can be cleanly separated instead of overloading one server.
- KubePlus's docs/example set (README, Operator-FAQ, getting-started, multitenancy
examples, etc.) is small enough to index locally without an external embedding API,
making it usable in offline/minikube classroom and CI environments.
- This gives students and adopters an agent that can answer "how do I configure X in
KubePlus" questions with citations back to the actual repo files, not just
general LLM knowledge.

### Proposed approach
1. **Ingest** (`ingest.py`): clone `cloud-ark/kubeplus`, walk `.md` docs and
`.yaml`/`.yml` example manifests (skip `vendor/`), chunk (~250 words, 50-word
overlap), fit a `TfidfVectorizer`, and persist to `index.pkl`.
- No embedding model download required — keeps it usable offline in minikube.
2. **Serve** (`rag_server.py`): a FastMCP server exposing one tool,
`search_kubeplus_docs(query, top_k)`, doing cosine-similarity lookup over the
saved index and returning `{source, score, text}` per match.
3. **Wire into kagent**: register the server as its own `ToolServer` CR and attach it
to the `Agent` alongside the existing tools `ToolServer`, e.g.:
```yaml
apiVersion: kagent.dev/v1alpha1
kind: ToolServer
metadata:
name: kubeplus-rag
spec:
stdio:
command: python
args: ["rag_server.py"]
workingDir: /path/to/kubeplus_rag
---
apiVersion: kagent.dev/v1alpha1
kind: Agent
metadata:
name: kubeplus-agent
spec:
tools:
- toolServer: your-existing-tools-server
- toolServer: kubeplus-rag
```

### Tasks
- [ ] Add `ingest.py` / `rag_server.py` (prototype attached/linked below)
- [ ] Decide on re-index trigger: manual re-run vs. scheduled job vs. git webhook
- [ ] Evaluate retrieval quality on a fixed set of test questions (multitenancy,
cross-ns-deps, storage isolation, etc.)
- [ ] Document the two-`ToolServer` pattern in the kagent + KubePlus integration docs
- [ ] (Stretch) swap TF-IDF for real embeddings (local sentence-transformers or an
embedding API) once past prototype stage

### Acceptance criteria
- A kagent `Agent` can be configured with both the existing tools `ToolServer` and
`kubeplus-rag`, and correctly calls `search_kubeplus_docs` when asked
KubePlus-specific questions.
- Retrieval returns relevant chunks (source file + score) for a documented set of
sample queries.
- Setup works fully offline against a local minikube cluster, with no external API
dependency for embeddings.

### Related
- Builds on the existing KubePlus MCP server work (kubectl-metrics wrapper,
kubeconfig security boundary, init-container plugin fetching).
- Complements the multi-tenant kagent isolation issues (cross-ns-deps annotation,
Agent Sandbox interoperability).

Start by reviewing the existing KubePlus MCP server and the proposed ToolServer and Agent YAML. Then examine the prototype for ingest.py and rag_server.py, and assess indexing the named Markdown and YAML sources offline. Done means both ToolServers can be attached to an Agent, search_kubeplus_docs returns relevant sources and scores for fixed sample queries, and the integration pattern is documented.

Written by the indexing model from the issue text.