Streaming resampling, shared by the engine and the entrypoints.
Lives here rather than beside the OpenAI audio mixin because the duplex engine needs it too, and engine code must not import the entrypoints layer. Pure numpy on purpose: importing this must not pull in torch or the API server.
Stateful polyphase resampler for streaming mono float audio.
Retains filter state so output is invariant to input chunk boundaries.