Enable JSON-schema based structured output by default

[SteveSandersonMS]

Update Following discussion, this change has been reduced to:

• Changing the default for useJsonSchema to true
• Renaming ModelId/Dimensions to DefaultModelId/DefaultModelDimensions

Plus other minor related renames and integration test fixes.

---

Old description

Fixes #6126 based on the design discussed there.

This introduces a new concept of per-model metadata. Usage:

var providerMetadata = chatClient.GetService<ChatClientMetadata>();
Console.WriteLine("Provider name: " + providerMetadata.ProviderName);

var defaultModelMetadata  = await providerMetadata.GetModelMetadataAsync();
Console.WriteLine("Supports native JSON schema: " + defaultModelMetadata.SupportsNativeJsonSchema);

var anotherModelMetadata = await providerMetadata.GetModelMetadataAsync("anotherModelId");

There's also the equivalent for embedding generators, and that's now how you can find the Dimensions value.

Structured output

GetResponse<T> still allows the developer to specify useNativeJsonSchema, but if it's unspecified, the default is now taken from the model metadata.

Providers

• Ollama always supports structured output
• OpenAI and Azure AI Inference vary based on the model

This PR bakes in knowledge about which model IDs currently support structured output. I'm on the fence about how much we really want to get into this business, but there's no way to get the info programmatically. Neither OpenAI or Azure AI Inference expose this metadata on the endpoints they have for describing models.

I've tried to make the logic as forgiving as possible if the external reality changes and new models appear, but having correct metadata here is a best-effort thing that developers would still have to override on a per-call basis if our default doesn't match their requirement.

The main alternative would be to remove the built-in knowledge of specific models and just default to "unknown" for the OpenAI/AIInference providers across all models. In effect that would retain the same behavior we have before this PR. However it doesn't move us forwards and cements an old-fashioned view of the world.

Yet another alternative would be defaulting to "true" for all models, since most OpenAI models do support it. However even some new-ish OpenAI models (e.g., o1-mini) don't support it, and of course many Azure AI Inference ones don't since it includes lots of small models.

Microsoft Reviewers: Open in CodeFlow

[shyamnamboodiripad]

Approving for the change under the Eval project 👍🏾

[luisquintanilla]

cc: @azchohfi @nmetulev for awareness and had similar asks as part of #5941