Keep normalization results floating point for integer inputs

[catplotlib]

Changes proposed in this pull request:

• Give the normalization primitives in gr00t/data/utils.py a floating-point result dtype, so integer inputs are no longer silently truncated.
• Add tests/gr00t/data/test_normalization_values.py, the first direct coverage for these four functions.

The bug

normalize_values_minmax, normalize_values_meanstd and unnormalize_values_meanstd size their output with np.zeros_like(values), which inherits an integer input dtype. The division is then truncated on assignment, so an integer state or action array is silently corrupted rather than normalized:

import numpy as np
from gr00t.data.utils import normalize_values_minmax, normalize_values_meanstd

values = np.array([[5, 7]], dtype=np.int64)

print(normalize_values_minmax(values, {"min": np.array([0.0, 0.0]), "max": np.array([10.0, 10.0])}))
# [[-1 -1]]          expected [[0.0, 0.4]]

print(normalize_values_meanstd(values, {"mean": np.array([4.0, 4.0]), "std": np.array([3.0, 3.0])}))
# [[0 1]]            expected [[0.333, 1.0]]

For min/max every value collapses to -1, because the truncated 0 then goes through 2 * 0 - 1. There is no warning; training or inference would just consume wrong numbers.

Passing the same values as float64 gives the correct results, so the behaviour depends entirely on the caller's dtype. StateActionProcessor forwards state[joint_group] and action[joint_group] straight into these functions without casting, and only one dataset path (sharded_single_step_dataset.py) casts to float32 on the way in.

To be clear about severity: every dataset shipped in demo_data/ declares dtype=float32, so I do not believe this is reachable through the datasets in this repo today. It is a latent trap for anyone feeding integer-valued state or action features — discrete grippers, mode flags, indices.

The fix uses np.result_type(values, np.float32), which keeps float32 inputs as float32 (no upcast on the common path) and promotes integers to float64.

Tests

normalize_values_* had no direct tests, so the new file covers the primitives rather than only the regression:

• min/max and mean/std normalize → unnormalize roundtrips
• the documented degenerate cases (min == max → 0, std == 0 → pass through)
• per-step 2D bounds and batched (B, T, D) inputs
• result dtype: integer inputs not truncated, float32/float64 preserved

pytest tests/gr00t/data/test_normalization_values.py -q

14 passed with this change. With the utils.py change reverted, the 3 dtype tests fail and the other 11 still pass, so the new coverage documents existing behaviour and only the regressions are new.

Wider suite on this branch: pytest tests/gr00t/data/ -m "not gpu" gives 213 passed. The 3 failures in test_dataset_factory.py and the collection error in test_embodiment_tags.py are pre-existing — they reproduce identically with this change stashed, and both come from an unrelated non-default argument 'diffusion_model_cfg' follows default argument error in my environment.

No new dependencies: the tests use seeded numpy rather than adding a property-testing library.

Before submitting

• I've read and followed all steps in the Making a pull request section of the CONTRIBUTING docs.
• I've updated or added any relevant docstrings.
• If this PR fixes a bug, I've added a test that will fail without my fix.
• If this PR adds a new feature, I've added tests that sufficiently cover my new functionality.

[catplotlib]

Bumping this one — CI is green. Anything I can do to help move it along, or anyone I should tag?

[catplotlib]

Following up once more on this one. CI is green and the change is contained to the normalization helper, so it should be cheap to review.

If it isn't a priority or the approach isn't what you want, that's completely fine - just say and I'll close it.