Support for downsampling output points or reducing per-point feature dimension in Sonata encoder

Hi there! I’m using the Sonata encoder to extract per-point features from a 4096‐point cloud, which gives me a [4096, 1088] feature tensor which is a lot.
I want to use those features to train another model and I only need a small subset of points (e.g. 1–50) or a much lower‐dimensional feature (e.g. <256) per point. Is there a built‐in way to either downsample to a fixed smaller number of points while preserving their features or to reduce the per‐point feature dimension in Sonata?
Thanks!

Start by locating the Sonata encoder entry point that produces the [4096, 1088] per-point feature tensor. Determine whether the requested fixed-point downsampling and lower per-point dimension are supported, and clarify the expected behavior for both options; done should provide a documented, usable path for producing the requested smaller outputs.

Written by the indexing model from the issue text.