VLM support for FoundationModels protocol

[carinapeng]

Purpose

Adds CoreAIVisionLanguageModel, a FoundationModels LanguageModel adapter for VLM bundles. This lets callers drive a vision-language model through a standard LanguageModelSession and pass images via the FoundationModels attachment API, the same surface used for first-party multimodal models.

Changes

• New swift/Sources/CoreAILanguageModels/VLM/CoreAIVisionLanguageModel.swift:
  • CoreAIVisionLanguageModel: LanguageModel — loads a VLM bundle folder, declares the .vision capability, builds the backing engine.
  • CoreAIVLMExecutor: LanguageModelExecutor — pulls the image attachment + text from the Transcript, runs the vision encoder + decode via the existing CoreAISequentialVLMEngine, and streams tokens into the FM channel.
• Thin adapter: reuses the existing engine (no export or runner changes).
• Streaming (session.streamResponse { … }) works the same way. An image attachment is required; a text-only prompt throws unsupportedTranscriptContent.

API usage

import CoreAILanguageModels
import FoundationModels

let model = try await CoreAIVisionLanguageModel(resourcesAt: bundleURL)
let session = LanguageModelSession(model: model)

let response = try await session.respond(
    options: GenerationOptions(maximumResponseTokens: 256)
) {
    Attachment(cgImage)          // or Attachment(imageURL:) / CIImage / CVPixelBuffer
    "What is in this image?"
}
print(response.content)