PyTorch adds oneDNN LSTM primitive for 40% faster XPU inference

PyTorch's latest commit adds native oneDNN LSTM primitive support for XPU inference, replacing the previous per-timestep fused cell approach. By using `dnnl::lstm_forward`, the entire LSTM sequence is processed in a single internal operation, eliminating O(T) kernel launches. The new path handles weight layout transformation from PyTorch's `[4*H, I]` to oneDNN's `ldgoi` format, weight reordering, and scratchpad management. Inference benchmarks on Intel PVC (Ponte Vecchio) with a bidirectional LSTM (hidden=256) show LSTM latency dropping from ~34ms to ~5ms and end-to-end Kokoro TTS inference improving by 40% (1.06s down to 0.635s). The implementation also includes packed sequence unwrapping for uniform batch sizes (common in batch=1 inference) and an XPU-specific LSTMCell fallback for training.

The change preserves correctness: gate order (i, f, c̃, o) is identical between oneDNN and PyTorch, and the primitive’s formula `gates = W·x + U·h + B` matches PyTorch’s approach with summed biases. Verified maximum differences are below 1e-6 compared to CPU reference across multiple configurations. The update only affects XPU hardware; CUDA and CPU inference paths remain unchanged. This is a significant performance boost for any LSTM‑based X PU deployment, from speech synthesis to NLP, and aligns PyTorch’s XPU backend with optimizations already available on other accelerators.