Lazy runner design: defer engine load

[carinapeng]

Purpose

Introduces lazy inference-engine lifecycle management for the FoundationModels adoption layer. Instead of CoreAILanguageModel holding a live engine from construction, engine load/unload is now owned by a new ModelResources type: lazy by default, loaded on first respond(...) (or explicitly via load()), borrowed per-generation, and releasable via unload(). This lets callers construct a model cheaply, query its on-disk footprint, and control when the (expensive) engine actually loads.

Changes

• ModelResources (new) — owns the engine's load/unload lifecycle. Concurrent callers share a single in-flight load; withEngine { } borrows the engine for the duration of a generation and defers teardown if an unload() races an active borrow. Shared per-configuration via a weak registry.
• CoreAILanguageModel — now builds directly from LanguageBundle (removes the CoreAIRunner.makeLanguageModel() hop). Adds:
  • LoadMode { .lazy, .eager } (default .lazy; .eager loads tokenizer + engine concurrently)
  • load() / unload() / estimatedSizeOnDiskBytes
  • guided-generation capability gating that works before the engine is resident (isGuidedGenerationSupported)
  • reasoning models default to a higher max-token budget
• prewarm — now kicks off the engine load in the background (non-blocking) instead of a synchronous semaphore wait.
• ResourceManaging — moved from CoreAIDiffusionPipeline/Components/ to CoreAIShared/Runtime/; diffusion pipelines updated to import it from the shared module.
• FileSize (new) — URL.recursiveFileSizeInBytes() helper backing estimatedSizeOnDiskBytes.
• Tests — ModelResourcesTests (load/unload/borrow races via an injectable loader), FileSizeTests, and PublicInterfaceTests additions.

[carinapeng]

CoreAILanguageModel.swift syncing with recent PRs

• Use custom AsyncSequences and avoid unstructured concurrency #24 try await strategy.decode → present
• 02a8edd Gemma additional-EOS → reconciled (additionalEosTokenIds)
• Fix pipeline race condition: rotate all buffers by pipeline depth #53 → doc comment reconciled
• Enable multi-turn KV cache reuse #64 remove engine.reset() → reconciled
• Make generate() async to serialize back-to-back turns #80 try await engine.generate → present