Add LiteRT-LM LoRA C API support

[leehack]

Summary

• add a build-time upstream patch that exports litert_lm_session_config_set_lora_file and litert_lm_session_config_set_audio_lora_file
• wire scoped text LoRA files through resource management into the LiteRT compiled-model executor
• preserve active LoRA IDs across context clone/restore and inject active LoRA buffers into decode inputs

Why

LiteRT-LM v0.13.1 has internal LoRA manager/session-config support, but the public C wrapper and text executor path do not expose enough behavior for Dart FFI clients. This enables the native runtime bundle needed by llamadart to support one LiteRT-LM LoRA adapter at scale 1.0.

Companion llamadart PR: leehack/llamadart#206

Validation

• python3 -m py_compile tools/build_upstream_runtime.py
• python3 tools/build_upstream_runtime.py --upstream-tag v0.13.1 --platform macos --arch arm64 --jobs 8
• nm -gU bin/macos/arm64/libLiteRtLm.dylib | rg 'litert_lm_session_config_set_(audio_)?lora_file|litert_lm_conversation_create'
• git diff --check

Runtime smoke used by companion PR

With this locally built macOS runtime, llamadart generated successfully through both ModelParams.loras and engine.setLora(...) using:

• model: /private/tmp/litert-lm-v0.13.1-src/runtime/testdata/test_lm.litertlm
• adapter: /private/tmp/litert-lm-v0.13.1-src/runtime/testdata/test_lora_rank32_f16_all_ones.tflite