ML Intern: run the Hub MCP server on the intern bouquet (#2570)
* ML Intern: run the Hub MCP server on the intern bouquet
Mode conversations now get https://huggingface.co/mcp?login&bouquet=intern
(hf-mcp-server >= 0.4.18) as their Hugging Face MCP entry. `login` stays so
the user's token is still forwarded and anonymous callers are refused;
`bouquet=intern` picks the Hub's tool preset for the mode. Ordinary chats
keep the env-configured `?login` entry.
The login-endpoint check now lives once in src/lib/utils/hf.ts and accepts
`login` plus at most a `bouquet` key. Anything else still fails closed:
`gradio=` in particular makes the Hub server hand the forwarded token to a
private Space named there.
A pinned Hub entry (own Authorization header) still wins the preset merge,
but gets `bouquet=intern` added unless the operator chose a bouquet.
The intern bouquet has no hub_repo_search, so the research sub-agent's
prompt only advertises it when the run has it, and the main prompt's
paper-search warning names hf_fs search instead.
* Research prompt: name the Spaces search scope in the no-repo-search fallback
