feat: Add fips-agents add vision (closes #20)

[rdwj]

Summary

Wires multimodal (image input) example client into existing agent
projects. Closes #20.

The new `add vision` command:

1. Verifies `server.files.enabled: true` in `agent.yaml` (precondition;
   the `file_id:` URL scheme resolves bytes via the BytesStore,
   which only exists when files is enabled). Refuses to apply with an
   actionable hint when not satisfied.
2. Drops `examples/vision_client.py` showing the three `image_url` URL
   forms the agent runtime accepts: inline `data:`, remote `https://`,
   and internal `file_id:`.
3. Prints next-steps describing how to point the agent at a
   vision-capable model (Granite Vision 3.2-2B canonical example) and
   run the example script.

Pairs with `fipsagents 0.20.0` (image input in OpenAI content blocks)
and `fips-agents add files`.

Notes on the issue wording

• No `vision:` section added to `agent.yaml`. The agent-template
  audit for #101 (image input on the runtime side) explicitly chose a
  single multimodal endpoint via the existing `model.endpoint` —
  adding a `vision:` block now would bake in a split that hasn't
  been needed yet.
• Example code lives at `examples/vision_client.py` (client-side),
  not `src/agent.py`. Content blocks are constructed by callers; the
  agent runtime resolves them automatically. The agent code itself
  doesn't need to change for image input.

Test plan

• `pytest` clean — 265 passed (was 261; +4 new tests for vision E2E
  + spec shape).
• `black --check` and `ruff check` clean on the changed Python files.
• Manual: `fips-agents add vision` in a fresh agent project
  without files enabled fails with the right hint; after
  `fips-agents add files`, succeeds and drops the example.
• Idempotent: second run of `fips-agents add vision` reports
  `already exists` and exits 0.