Add recurrent SNN docs, tests, and benchmark scaffold

[Ikaikaalika]

Summary

• add a recurrent-temporal usage section to the NN docs with a chunked streaming pattern
• add recurrent regression tests for dtype propagation, gradient parity, and long-sequence numerical finiteness
• add a benchmark scaffold comparing unrolled tanh RNN and LIF hard-reset workloads across configurable shapes/dtypes

Validation

• python -m pytest python/tests/test_nn.py -k "recurrent_dtype_propagation or recurrent_gradient_parity or recurrent_long_sequence_stability" -q
• python benchmarks/python/recurrent_snn_bench.py --batch-sizes 1 --sequence-lengths 8 --hidden-sizes 16 --input-size 8 --dtypes float32 --warmup 1 --iters 2

Notes

• benchmark sample output artifacts under benchmarks/python/results/ were intentionally left uncommitted