feat(vlm): add Gemma3n audio reference path (#875)

[inureyes]

Summary

• Implement the pinned Gemma3n waveform processor, audio Conformer tower, projection/pooling, hard/soft audio embedding merge, and per-layer-input integration.
• Integrate audio placeholders and preparation into CLI and OpenAI-compatible serving while keeping request-scoped multimodal state isolated across concurrent sequences.
• Match the maintained reference's left-padding contract, BF16 convolution boundary, and finite dtype-specific attention mask minimum required for token-exact greedy ASR.

Validation

• Pinned Transformers source: 181beb3ba4c47098ed8cbc97ee250d1d45ae0107; pinned checkpoint revision: af70430f84ea4d7ac191aaa2bd8e14d2a5e8f6ee.
• GB10 CUDA CLI generated the official 16-token sequence exactly: Chapter 1. Mrs. Rachel Lind is surprised.\n\nMrs. Rachel Lind.
• OpenAI-compatible streaming generated the same 16-token sequence across SSE deltas and completed with [DONE].
• cargo test --release --features cuda gemma3n --lib -- --test-threads=1: 34 passed, 0 failed, 14 ignored.
• cargo test --release --features cuda gemma3n --lib -- --ignored --test-threads=1: 14 passed, 0 failed, including request-scoped per-layer-input isolation tests.
• cargo fmt --all -- --check, cargo check --release --all-targets --features cuda, and cargo clippy --release --lib --bins --tests --features cuda -- -D warnings passed with Rust 1.93.1.
• The stricter all-target Clippy invocation additionally reaches a pre-existing print_literal warning in examples/qmm_gemv_microbench.rs that is unchanged from origin/main; all issue-attributable library, binary, and test targets are clean.

Closes #875