pub fn draw_batch(
input: PolyaGammaBatchInput<'_>,
) -> Result<Array1<f64>, String>Expand description
Top-level entry point: dispatches to GPU when enabled, available, and
admitted by the calibrated fused-batch crossover; otherwise CPU.
Both paths are deterministic for a fixed seed. The CPU path delegates to
the upstream sampler while the CUDA path is independently validated against
it in distribution. CUDA probe and execution faults are returned; only a
size-policy refusal or lossless Ok(None) availability result selects the
CPU implementation.