Upgrade Metal backend to unified GPU/ANE architecture

[ChinChangYang]

Summary

• Replace separate CoreML and Metal backends with a single unified Metal backend based on lightvector/KataGo#1148, enabling per-thread GPU/ANE multiplexing via the vendored katagocoreml C++ library for on-the-fly .bin.gz → CoreML model conversion
• Change the built-in model from a pre-converted CoreML .mlpackage (b28c512) to an 18-block .bin.gz network that the Metal backend converts at runtime, eliminating the need for pre-converted CoreML models
• Simplify the iOS app interface so all models (built-in and downloaded) use the same Metal backend path, with iOS defaulting to ANE (Neural Engine) for power efficiency and macOS defaulting to GPU (MPSGraph)

Key changes

C++ Backend

• metalbackend.cpp/h — unified backend with CoreMLConversion namespace for native model conversion, per-thread METAL_MUX_GPU(0) / METAL_MUX_ANE(100) multiplexer
• metalbackend.swift — rewritten for CoreML model loading/inference via CoreMLComputeHandle
• metallayers.swift — new file with MPSGraph GPU layer implementations and MPSGraphModelHandle
• cpp/external/katagocoreml/ — vendored C++ library (88 files) for .bin.gz → .mlpackage conversion
• Removed coremlbackend.cpp/h/swift, coremlmodel.swift, and all CoreML-specific code paths in nneval, setup, gtp, commandline

iOS App

• KataGoHelper.swift — simplified: removed coremlModelPath/useMetal params; uses metalDeviceToUseThread0 + metalUseFP16
• ModelRunnerView.swift — unified model loading (no more builtIn vs downloaded backend distinction)
• NeuralNetworkModel.swift — built-in model is now "Built-in KataGo 18b Network" (.bin.gz)
• default_gtp.cfg — Metal backend config (metalDeviceToUseThread0=100, metalUseFP16=true)

CI/Build

• ci_post_clone.sh — installs cmake/ninja/protobuf/abseil, builds katagocoreml for all platforms, fetches 18b model
• build.yml — removed CoreML .mlpackage download/test steps
• Xcode project — added metallayers.swift, removed CoreML files, USE_COREML_BACKEND → USE_METAL_BACKEND, added katagocoreml linker flags

Test plan

• Build for iOS Simulator, macOS, and visionOS Simulator
• Verify built-in 18b model loads and converts to CoreML on first launch
• Verify downloaded .bin.gz models work through the Metal backend
• Test ANE inference on iOS device
• Test GPU inference on macOS
• Verify human SL model still loads correctly
• Run Xcode Cloud CI pipeline end-to-end

https://claude.ai/code/session_01LLDk6NpbEpchikWcq21ksf