v0.7.0

• Add Python bindings: pip install speakrs ships one abi3 wheel per platform, with CPU and CoreML on macOS arm64 and CPU and CUDA on Linux x86_64, plus a speakrs diarize command and a speaker-labeled transcript example; see crates/speakrs-python
• Add speakrs::models::{MODEL_REPOSITORY, MODEL_REVISION, required_model_files}, available without the online feature, so callers can download or verify the files a mode loads with their own HTTP client
• Add QueuedDiarizationJobId::as_u64
• Add QueueReceiver::recv_timeout, so a caller can wait for results in bounded slices and stay responsive to interrupts
• Fix CUDA builds embedding every PTX and cubin file more than once, which roughly doubled the size of the linked binary
• Run the RTX 5090 embedding trunk on FP16 tensor cores: a built-in recipe moves each layer to FP16 above a per-layer batch size, sm120 kernels pair taps into m16n8k16 products, and FP16 Winograd kernels run the 64-, 128-, and 256-channel layers. On VoxConverse dev, cuda drops from about 39 s to 25 s with DER moving from 7.0123 to 7.0119, and cuda-fast from about 20 s to 13 s with DER unchanged; other GPUs keep their kernels
• Keep the GPU busy between CUDA batches and files: the next embedding batch launches before the current one is read back, and in concurrent mode the next file's segmentation runs during the current file's last embedding batches; results and their order don't change
• Speed up the tensor-core segmentation convolutions by 0.7 to 1.2% on an RTX 5090, with bitwise-identical output
• Drop zip from the dependency tree: ndarray-npy no longer enables its npz feature, which pulled in zip 6 with RUSTSEC-2026-0367, RUSTSEC-2026-0368, and RUSTSEC-2026-0369
• Update the locked rustls-webpki to 0.103.16 and webpki-roots and webpki-root-certs to 1.0.10 for RUSTSEC-2026-0381, RUSTSEC-2026-0382, and RUSTSEC-2026-0383