fix(tts): cast bf16 tables explicitly where the CUDA overlay demotes

[inureyes]

Summary

Five of the six CUDA lib-test failures share one cause: mlxcel's own CUDA overlay of MLX's promotion table (src/lib/mlx-cpp/patches-cuda/dtype.cpp) resolves bf16 with f32 to bf16 on every device of a CUDA build. It is not an MLX bug at the pin and not an mlxcel-core wrapper. The overlay is deliberate (#636, single-dtype bf16 decode graph) and stays, so the code that needed f32 now casts explicitly. The sixth failure (gelu) was already fixed on main by #2080 and passes on CUDA unchanged.

• Production fix (TTS): RvqCodebooks::depthsum_embedding and MogHead::infer gathered rows from stored bf16 tables and combined them with f32, so on CUDA the depth-sum and the MoG mu came back bf16. Both now cast the gathered rows (exact, bit-identical off CUDA). Reverting only these casts makes depthsum_ignores_mask_index and mog_infer_shapes_are_finite_with_guidance fail on CUDA again.
• Probe test mixed_bf16_f32_promotion_follows_the_build pins add, matmul and fast_rms_norm dtype for f32 against bf16: f32 without cuda, bf16 with it. The issue asked for f32 on both builds; that cannot hold while the overlay is kept, so the probe pins the contract that does hold.
• Tests: the f32_weights and perception references spell upstream promotion out as an in-graph astype on CUDA; the MoG test loads weights through promoted_subset as RvqEarTtsModel does.
• Technical report EN+KO added.

Test plan

• GB10 --release --features cuda, under gpu-lock, --test-threads=1: audio::f32_weights 4, audio::fastconformer 14, models::nemotron_voicechat::tts 13, models::gemma3_backbone 5, all pass
• CPU-only Linux build (no GPU feature, separate target dir): same filters, same counts, all pass. Not MLXCEL_DEVICE=cpu, since the overlay is compiled in.
• cargo clippy --release --features cuda --lib --tests -- -D warnings and cargo fmt --check clean

Not validated here:

• The plain Linux CPU lib-test build does not link on main: turbo/kv_inplace_write.cpp (perf(cache): write the decode KV row in place instead of copying the cache #1961) references mlx::core::copy_gpu_inplace, undefined without a GPU backend. The CPU run used a linker wrapper ignoring that one unresolved symbol.
• The EAR-TTS backbone still meets f32 activations with stored bf16 tensors (Gemma norm weights, bos_emb, null_emb, audio_prompt_projection_w), so by the probed rule it runs bf16 on CUDA where the reference runs f32. Not measured end to end; no VoiceChat checkpoint on this host.

Closes #2087