Context hpc datasets

[Samuel-WM]

This PR adds the initial end-to-end HPC/DDP implementation for the sklearn-style interface.

Before opening the PR, I ran scaling-oriented validation on the new distributed path to confirm the training loop was stable under multi-rank execution and that performance characteristics (epoch time / throughput / step time) behaved as expected as world size increased. Those checks focused on fit and predict workloads and included basic correctness signals.

After those initial runs, I discovered an implementation mistake introduced during subsequent development on the DDP prediction/aggregation. The reason this was not obvious from the earlier scaling results is that DDP training stability and scaling can look correct even when the predict path is missing the required cross-rank aggregation details (synchronizing outputs, restoring global row ordering, and deduplicating sampler padding). I have made the necessary corrections in this branch, and am redoing the benchmark and perturbation tests with the updated package.

[cnellington]

See if you can get the tests to run. If not, move to a branch on the repo instead of your fork.