Add Foundry Local as a chat and embedding provider in the AI Chat Web template

[luisquintanilla]

Summary

The AI Chat Web template (dotnet new aichatweb, Microsoft.Extensions.AI.Templates) lets
you pick a chat and embedding provider: Azure OpenAI, GitHub Models, OpenAI, and Ollama. The
only on-device option today is Ollama, a third-party runtime. This proposes adding
Foundry Local, Microsoft's own on-device inference runtime, as
a first-class provider for both the chat IChatClient and the embedding IEmbeddingGenerator.

Why

• It gives the template a Microsoft-owned local inference story: no cloud, no API key, models
  run on the developer's machine on CPU, GPU, or NPU through ONNX Runtime.
• Foundry Local is OpenAI-compatible, so it drops into the same
  OpenAIClient -> AsIChatClient() / AsIEmbeddingGenerator() shape the template already uses
  for the OpenAI and GitHub Models providers. The change is small and well-contained.
• The template's provider list is the de facto answer to "what does .NET AI support locally."
  Adding Foundry Local closes the gap and gives the docs and RAG tutorials a reference
  integration to point at.

What it would look like

A new provider choice, for example --provider foundrylocal, that wires Foundry Local for
both chat and embeddings. Suggested defaults: qwen3-4b for chat (Apache-2.0, tool-calling
capable) and qwen3-embedding-0.6b for embeddings (1024-dimension vectors).

Proof of concept

A runnable end-to-end sample is here:
https://github.com/luisquintanilla/foundry-local-aichatweb

It is the standard AI Chat Web app (Blazor, RAG over a local PDF, SQLite vector store) with the
provider swapped to Foundry Local. It builds and runs end to end, and includes a dev container
so you can try it in Codespaces.

An aspire branch shows what an Aspire orchestration could look like, using a small custom
AddFoundryLocal hosting integration (there is no Aspire hosting integration for Foundry Local
today): https://github.com/luisquintanilla/foundry-local-aichatweb/tree/aspire

Notes for implementation

• SDK: Microsoft.AI.Foundry.Local 1.2.3 (stable, on nuget.org). The app uses the manager to
  start the local OpenAI-compatible web service and load models, then points the standard
  OpenAI client at it. Foundry Local is keyless on localhost.
• The embedding model returns 1024-dimension vectors, so IngestedChunk.VectorDimensions
  (currently hardcoded to 1536) needs to become per-provider.
• First run downloads the model weights; later runs start fast.

Scope

Non-Aspire first. The Aspire path is under investigation: there is no Aspire hosting
integration for Foundry Local today (unlike Ollama's AddOllama), so a custom
AddFoundryLocal integration is prototyped on the sample's aspire branch (linked above). We
can decide how to bring Aspire into the template once that settles.