Develop versatile input/output management for indefinite-sized outputs in ExKernel

Develop a more versatile input/output management mechanism within the ExKernel programming model to safely and efficiently support operators that produce indefinite-sized outputs while operating on datasets that exceed GPU memory capacity.

Background

The ExKernel interface in Vortex maps inputs and outputs into chunks that fit GPU memory and allows kernels to operate in a pipelined fashion while Exchange handles IO. For operators with indefinite output sizes, the current guidance is to conservatively size input chunks to avoid GPU memory overflow, which can be suboptimal.

A more robust input/output management mechanism is needed to handle dynamic or unknown output sizes without over-constraining chunk sizes, improving both usability and performance for a wider range of operators.