fix: Nemotron VoiceChat TTS backbone runs bf16 on CUDA #2109

Problem

The CUDA overlay src/lib/mlx-cpp/patches-cuda/dtype.cpp promotes bf16+f32 to bf16 on purpose (#636). #2093 made the code embeddings (tts/rvq.rs) and MoG head (tts/mog_head.rs) cast explicitly, but in src/models/nemotron_voicechat/tts/model.rs three bf16 tensors still meet the f32 backbone stream, so on CUDA they demote it while the reference runs f32:

1. bos_emb and null_emb (loaded at :124-125, used at :298 multiply and :220 broadcast_to).
2. audio_prompt_projection_W (loaded at :149, matmul at :295).
3. The backbone's GemmaRMSNorm (src/lib/mlxcel-core/src/layers.rs:588-590, used via src/models/gemma3_backbone.rs:158 and per layer) builds 1 + w in the stored bf16 dtype. The CUDA rms_norm (patches-cuda/fast.cpp:98-140) takes out_type = result_type(x, weight), so it casts the f32 input down to bf16. The TTS-local OffsetRmsNorm (tts/norm_mlp.rs:93-95) already builds in f32 and is fine.

All three are reasoned from the op rules, not measured: there is no VoiceChat checkpoint on GB10.

Proposed fix

• Add bos_emb, null_emb and audio_prompt_projection_W to promotes_to_f32 (model.rs:77); widening them is exact.
• Do not add the norm keys: promoting w before 1 + w changes rounding on Metal and CPU. In the VoiceChat load path, widen the built bf16 1 + w to f32 instead. GemmaRMSNorm is shared by every Gemma model, so leave layers.rs unchanged.
• Update f32_promotion_keeps_norms_and_the_subword_path_as_stored (tts/tts_tests.rs:443), which currently asserts that null_emb is not promoted.

Acceptance Criteria

• A new synthetic-weight test (bf16 weights) asserts that the backbone hidden stream is f32 after embedding assembly, after a backbone norm and after the projection. Under --features cuda it fails on 33c4505.
• On non-CUDA builds, TTS output for a fixed seed is bit-identical before and after the fix. (argued from the code in PR fix: keep the VoiceChat TTS backbone stream f32 on CUDA #2115, not measured: the CPU-only build is blocked by fix: CPU-only Linux build fails to link kv_inplace_write #2108)
• The PR states that no real checkpoint was used for validation.

Do it in one PR, run the checks once at the end, and push as few times as possible.

Verification

gpu-lock run --tag voicechat-dtype -- cargo test --release --features cuda --lib nemotron_voicechat -- --test-threads=1