Polars assisted sampling for KumoRelational

[alan-kumo]

"add new example to KumoRelational ... using Polars assisted sampling" — requested workflow.

Adds a reusable Parquet-backed relational sampler and a self-contained rel-amazon user-churn example. The sampler retains relationship keys and timestamps in CPU memory, uses SDM's existing neighbor sampler, and reads selected feature rows from Parquet with Polars. The sampled RelatedTables feed directly into KumoRelational.

The example ran against 23,218,245 source rows on a 15 GiB RAM machine. With 512 context and 64 query task rows, it sampled 5,018 and 1,019 related rows, reached 4.95 GiB peak process RAM, and produced 0.7344 query accuracy. The focused sampler test checks exact sampled-table parity with the in-memory sampler; the example documents the remaining graph metadata held in RAM.

Validated with eight focused relational tests, the end-to-end rel-amazon run, and pre-commit run --all-files.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.