feat: start Qwen3-ASR support

[sxwxs]

Summary

Starts Qwen3-ASR support with the parts that do not depend on a new private NPU kernel ABI:

• add a C++ implementation of the Qwen3-ASR waveform frontend;
• add exact audio-token length calculation;
• add prompt construction and output parsing helpers;
• add standalone C++ tests and a Transformers numerical parity check.

Preprocessing parity

The test covers:

• 16 kHz mono float input;
• minimum 8000-sample padding;
• periodic Hann / centered reflect-padded STFT;
• 128-bin Slaney mel filter bank;
• Qwen log-mel normalization;
• right-padding to 2 * n_window;
• feature-major [mel_bins, frames] output layout.

Observed against the Transformers reference on the deterministic fixture:

shape: (128, 100)
max abs error: 3.4332275e-05
mean abs error: 2.960395e-07

Test

cd src/test/qwen3_asr_preprocessor
cmake -S . -B build -G Ninja
cmake --build build
ctest --test-dir build --output-on-failure
./build/test_qwen3_asr_preprocessor --dump /tmp/qwen3_asr_features.bin
python3 reference.py /tmp/qwen3_asr_features.bin

Remaining blocker

The existing public qwen3_npu interface accepts token IDs but does not document a payload for replacing audio placeholder embeddings. The next stage needs one of:

• the existing qwen3_npu::prefill(..., void* payload) payload contract;
• a mixed token/input-embedding prefill API;
• or a Qwen3-ASR-specific engine.

Related: #616.

[sxwxs]

Local validation update:

• Full cmake --preset linux-default && cmake --build build -j4 succeeds.
• The resulting FLM 0.9.46 binary runs an existing Qwen3.5-2B model on the XDNA2 NPU successfully.
• Standalone C++ tests and Transformers preprocessing parity still pass.

The remaining dependency for the next implementation stage is the mixed token/input-embedding prefill contract described in the PR.