LSTM and duration FP32 references

Invoke-KokoroBidirectionalLstm.ps1 implements one batch-one, single-layer bidirectional LSTM with PyTorch's i, f, g, o gate order, independent forward and reverse state, zero initial state, and concatenated output. Its behavior follows torch/nn/modules/rnn.py at 2b3ec34829036a65cd9d1398ea72a0167dc37470 and Kokoro's usage in kokoro/modules.py at dfb907a02bba8152ca444717ca5d78747ccb4bec. The reference is bounded and written in PowerShell; it is not a device implementation.

Invoke-KokoroDurationPrediction.ps1 connects the stock 512-to-50 duration projection to New-KokoroDurationMap.ps1, which performs sigmoid, bin sum, speed division, rounding, minimum-one clamp, and frame-to-token expansion as specified by pinned kokoro/model.py. This consumes an LSTM output; it does not yet compute the style-conditioned duration encoder that feeds the LSTM.

The pinned checkpoint digest was verified before tensor reads. Executable gates Test-KokoroBidirectionalLstm.ps1, Test-KokoroDurationPrediction.ps1, and Test-KokoroDurationMap.ps1 pass analytic cases, stock tensor shapes, and finite stock-weight output. These tests are not an independent stock numerical differential, a full duration prediction from phonemes, or a PCM/speaker result.