Add Gemma 3n E4B audio encoder (Conformer) support

12-layer USM-style Conformer audio encoder for Gemma 3n E4B multimodal models.
Enables on-device audio encoding: mel spectrogram → Conformer → embedder pipeline.

New files:
- Gemma3nAudio.swift: Full Conformer port (chunked local attention, depthwise conv1d,
  cumulative group norm, temporal reduction, sub-sample conv projection)
- Gemma3nAudioConfig.swift: 28 audio encoder configuration parameters
- Gemma3nVLM.swift: Top-level VLM wrapper with audio embedding injection
- Gemma3nAudioTests.swift: Configuration decoding tests

Architecture: mel [1,T,128] → SubSampleConv (4x) → 12 Conformer blocks →
temporal reduction (4x) → AudioEmbedder (1536→2048) → LM token stream

Tested on iPhone: 15.8s audio → 99 tokens in 0.48s

Author: vahsaechao