onnx: LSTM / GRU / RNN, unrolled over the static sequence length

[sbryngelson]

The last big compositional tier: the recurrent family, unrolled the same way Loop (#91) proved out — the whole recurrence becomes one fused program.

What

• LSTM — per timestep, one fused linear over the stacked input-side gate weights and one over the recurrent side (iofc order), sliced per gate. Forward, reverse, and bidirectional; Y, Y_h, Y_c outputs.
• GRU — zrh order, both linear_before_reset forms (the before-reset form reuses the fused recurrent matmul; the after-reset form takes one extra linear over the h-gate slice).
• RNN — vanilla tanh.
• Shared scope, rejected loudly: non-default activations, clip, LSTM peepholes, per-sample sequence_lens, layout=1, tensor-valued weights/initial states.
• Importer hardening these forced: ONNX marks omitted optional inputs and skipped outputs with empty-string names; "" inputs now resolve to None and "" outputs are no longer stored (previously they'd pollute the value table under the key "").
• Docs: table rows + scope note; also updates the two Gelu entries onnx: support tanh-approximate Gelu #92 made stale (exact + tanh-approximate now).

Verification (on-device, M5)

pytest tests/test_onnx.py: 166 passed — LSTM forward/bidirectional/Y_h, GRU in both reset forms, reverse RNN, all against onnxruntime at fp16 tolerance, plus the peephole reject. One test bug caught during rebase verification: the weight seeds used Python's salted hash(), making the numerics nondeterministic across processes — now zlib.crc32, verified stable across three consecutive runs. ruff, 2-space pylint, pyright (0 errors) clean.