Fix truncated-horizon evaluation masks

[Arison591]

Summary

• Slice action_is_pad to the requested action_horizon alongside the ground-truth action chunk.
• Add a focused regression test covering a shorter evaluation horizon with padded actions.

Root cause and impact

evaluate_batch() supports evaluating fewer action steps than are stored in the batch. It already sliced action_gt to that horizon, but passed the full-length padding mask to action_mse(). For example, evaluating a 2-step horizon from a 4-step chunk failed while expanding a [B, 4] mask over a [B, 2, D] tensor:

RuntimeError: The expanded size of the tensor (2) must match the existing size (4) at non-singleton dimension 1

The mask now follows the same horizon as the target, so truncated-horizon metrics both run and continue to exclude padded steps.

Validation

• python -m pytest -q tests/test_metrics.py — 1 passed
• python -m ruff check starwam/training/metrics.py tests/test_metrics.py — passed
• python -m compileall -q starwam tests — passed
• Repeated the regression on an NVIDIA GeForce RTX 5090 (SM 12.0, PyTorch 2.7.0+cu128) — returned the expected {"action_mse": 1.0}

No model weights, dataset, or external service are required for the regression test.