Add LiteRT trait and LiteRTLanguageModel backend

[john-rocky]

Implements #174: a LiteRT package trait and a LiteRTLanguageModel backend that runs .litertlm models (Gemma 4 E2B/E4B, or any .litertlm on the Hub) fully on-device via Google's LiteRT-LM runtime, with Metal GPU acceleration on iOS and macOS.

let model = LiteRTLanguageModel(model: .gemma4_E2B)
let session = LanguageModelSession(model: model)
let response = try await session.respond(to: "What is the capital of France?")

What's included

• Package.swift: a LiteRT trait; the john-rocky/swift-litert-lm dependency is gated by .when(platforms: [.iOS, .macOS], traits: ["LiteRT"]), so default builds and other platforms are unaffected.
• LiteRTLanguageModel (behind #if LiteRT), following the existing backend patterns:
  • Three initializers: catalog model (auto-download from the Hub), any Hugging Face repo hosting a .litertlm, or a local file. Engine bring-up is lazy (first request or prewarm), matching how MLXLanguageModel defers loading.
  • The session transcript maps to a LiteRT-LM conversation per request (system message from .instructions + prior turns as initialMessages + trigger message), so multi-turn sessions and multiple sessions per model instance behave correctly.
  • respond / streamResponse for text; image segments (data, file URL, or fetched remote URL) for models with a vision tower.
  • Structured generation via schema-in-prompt + balanced-JSON extraction (honors includeSchemaInPrompt).
  • Tool calling for respond: prompt-driven tool descriptions, tool-call parsing, and the standard ToolExecutionDecision delegate flow, with the result fed back for a final answer (capped at 4 rounds).
  • GenerationOptions mapping: temperature, .greedy, .topK, .nucleus → LiteRT-LM's sampler config.
• Tests: env-gated suite (LITERT_TEST_MODEL=/path/to/model.litertlm), same spirit as the MLX suite — the weights are multi-GB, so they don't run in CI.
• README: provider list, trait docs, image-support table, and a LiteRT-LM usage section.

Verification

• swift build --traits LiteRT — builds clean (no warnings) on macOS arm64.
• swift build (default traits) — unaffected; the new file is fully #if LiteRT-gated and the dependency isn't resolved into the build graph without the trait.
• swift format lint — clean on the new files.

Known limitations (documented in the README section)

• Tool calling is respond-only for now (same starting point as MLX, cf. MLX: support tool calling during streamResponse (currently respond()-only) #164).
• maximumResponseTokens isn't mapped (LiteRT-LM sets the token budget at engine bring-up, not per request).
• Audio understanding exists in the underlying runtime, but Transcript.Segment has no audio case yet, so this backend is text + image.