feat(vision): Gemma 4 audio input (gemma4ua / gemma4a)

[pekkah]

Follow-up to #250 / PR #251 (image input shipped). Gemma 4 is natively multimodal incl. audio, which is not implemented.

Weight-availability findings (from dumped mmproj headers, 2026-06-15)

• 12B QAT mmproj (mmproj-gemma-4-12b-it-qat-q4_0.gguf): clip.audio.projector_type=gemma4ua and has_audio_encoder=True are declared, but the file ships ZERO a.* audio tensors (only v.* vision + mm.*). So 12B audio is not possible with this QAT mmproj — it would need an audio-carrying 12B mmproj (verify one exists upstream).
• E4B mmproj (gemma-4-E4B-it-mmproj.gguf): ships a real audio encoder — clip.audio.projector_type=gemma4a, 751 a.* tensors (~12-block conformer-style: num_mel_bins=128, embedding 1024, ffn 4096, 8 heads), mm.a.input_projection [1536→2560].

Scope

• Audio preprocessing (mel-spectrogram, 128 mel bins) for the gemma4a path.
• gemma4a audio encoder forward pass (conformer blocks incl. norm_conv) — part of the E4B encoder-full effort.
• Project audio features → text embed dim and splice via the existing ForwardEmbedding seam at <|audio|> placeholders (<|audio>/<audio|> wrap).
• CLI --audio + (later) server audio content blocks.

Relationship

• The E4B audio encoder (gemma4a) is part of the E4B encoder-full work — see Add Gemma 4 E4B multimodal (vision / image input) support #126 (its deferred V6). This issue can either subsume that or stay the cross-cutting audio tracker.
• 12B audio (gemma4ua, encoder-free like its vision) is blocked on weight availability (above).