TorchTrainDataset loads more data than needed

[jder]

We are currently loading each step of an autoregressive rollout independently in training. This means that when we're not using dask (which is the usual case) we duplicate loads for adjacent steps. We should coalesce these across the full time range for one load, given hist etc.

@fomo-bot can you confirm the problem and propose a solution in a PR?

[jder]

Confirmed the problem: TorchTrainDataset.__getitem__ was selecting and materializing each autoregressive step independently. With shared input/label prognostic sources, the step-N label range is reused as the step-(N+1) input range, so non-dask loading can duplicate adjacent time reads.

Opened draft PR #799 with a proposed fix: #799

The PR coalesces each sample's shared prognostic rollout range into one selected tensor, then slices that tensor into per-step examples. Cross-resolution input/label sources still use separate selected ranges. Focused dataset tests, schedule coverage, ruff, and touched-file pre-commit pass; the full tests/test_datasets.py run is blocked later by missing flash_perceiver in this environment for unrelated samudra_multi inference setup.