docs(vertexai): replace incorrect model identifiers with auto-discovery note

[major]

What does this PR do?

The Vertex AI provider description listed model identifiers with a vertex_ai/ prefix (e.g. vertex_ai/gemini-2.5-flash) that never matched the resource names returned by the Vertex AI API (publishers/google/models/gemini-2.5-flash), causing server startup failures when users followed the documentation.

Since the provider rewrite in #4951, models are dynamically discovered from the Vertex AI project at startup via list_models(). The hardcoded model list was stale and misleading. This replaces it with a note that models are auto-discovered.

Closes #5403

How model discovery works after #4951

• Auto-discovery (default): list_models() queries the Vertex AI API at startup and registers all available models automatically. No user configuration needed. The routing table prefixes each model with the provider ID, so publishers/google/models/gemini-2.5-flash becomes vertexai/publishers/google/models/gemini-2.5-flash.
• Short-form model IDs work too: API requests can use vertexai/gemini-2.5-flash as a shorthand. The google-genai SDK normalizes the name internally.
• Restricting models via allowed_models: Users who want a subset can set allowed_models in the provider config using the full API resource names (e.g. publishers/google/models/gemini-2.5-flash).
• The old vertex_ai/gemini-2.5-flash format was never valid as a provider_resource_id and failed check_model_availability() against the API-populated cache.

Verified with a live Vertex AI project

Ran llama stack run with remote::vertexai against a real GCP project. Server auto-discovered 13 models at startup:

$ curl -s http://localhost:8321/v1/models | python3 -c "import sys,json; [print(m['id']) for m in json.load(sys.stdin)['data']]"
vertexai/publishers/google/models/gemini-1.5-pro-002
vertexai/publishers/google/models/gemini-2.0-flash-001
vertexai/publishers/google/models/gemini-2.0-flash
vertexai/publishers/google/models/gemini-2.0-flash-lite-001
vertexai/publishers/google/models/gemini-2.5-flash-preview-04-17
vertexai/publishers/google/models/gemini-2.5-pro-exp-03-25
vertexai/publishers/google/models/gemini-2.5-pro
vertexai/publishers/google/models/gemini-2.5-flash
vertexai/publishers/google/models/gemini-2.5-flash-lite
vertexai/publishers/google/models/gemini-2.5-pro-tts
vertexai/publishers/google/models/gemini-2.5-flash-tts
vertexai/publishers/google/models/gemini-live-2.5-flash-native-audio
vertexai/publishers/google/models/gemini-3.1-flash-image-preview

Chat completion with short-form ID works:

$ curl -s http://localhost:8321/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "vertexai/gemini-2.5-flash", "messages": [{"role": "user", "content": "Say hello in exactly 5 words."}]}' \
  | python3 -m json.tool
{
    "id": "chatcmpl-...",
    "choices": [{"message": {"role": "assistant", "content": "Hello there, my dear friend."}, "finish_reason": "stop", ...}],
    "model": "vertexai/gemini-2.5-flash",
    ...
}

Test Plan

Documentation-only change. Verified provider_codegen.py regenerates remote_vertexai.mdx cleanly and all pre-commit hooks pass.

$ uv run ./scripts/provider_codegen.py
Processing provider registry
  Processing provider registry...

$ SKIP=actionlint git commit ...
Provider Codegen.............................................................Passed
ruff (legacy alias)..........................................................Passed
ruff format..................................................................Passed
mypy.........................................................................Passed
markdownlint.............................................(no files to check)Skipped