v0.15.0

• Embed ordered text, images and audio with typed inputs and modality probes on
  native llama.cpp, native CPU LiteRT-LM (opt in with
  liteRtLmEmbeddingEngine) and compatible Web llama.cpp assets.
• Fix vision/audio projector and media-embedding loading in Windows
  dart build cli bundles, which also affected 0.14.
• Update the native llama.cpp runtime and Apple companion to v0.7.0-1;
  upgrade core and all Apple companions to 0.15.x together.

• Aligned the default WebGPU bridge assets to v0.1.65 for image/audio
  embeddings (API floor v0.1.64+), qualified on Chrome CPU wasm memory64
  against native CPU/Metal v0.7.0-1 with upstream
  v0.7.0@620d2225cc436ab0b001e3200e590e3b286ef9ab. Web LiteRT-LM remains
  @litert-lm/core@0.18.0. Immutable Web asset manifest:
  185cfc060fb90b7c8d240b23e7eeb85ef402eb3ef52e8acb1a5590cd5da56328.

• Behavior change: Run decision heads on CPU and report CPU with zero GPU
  layers when native llama.cpp single-device placement selects no GPU.
• Behavior change: Reject embedding, scoring, generation and LoRA calls on a
  context during native text-to-speech with LlamaStateException.
• Keep performance counters accurate across cancellation, validation and
  dispatch failures, reloads, embeddings, scoring and speech requests.
• Preserve LiteRT-LM Web runtime failures that race generation cancellation.
• Correct chat-template error labels for empty model templates.
• Probe native thinking-budget support and report missing speculative draft
  models as typed argument errors.
• Behavior change: Report WebGPU argument rejections as typed argument
  errors instead of inference failures.