test(data): fix flaky quantized depth roundtrip tests

[janickm]

Problem

The test_quantized_depth_roundtrip test in ncore/impl/data/v4/components_test.py fails intermittently in CI (e.g. run 27951090321):

test_quantized_depth_roundtrip
>   np.testing.assert_allclose(dequantized, original, atol=0.5 * quant.scale, rtol=0)
E   AssertionError

Root cause

Two issues compound:

1. Unseeded RNG -- the three quantized-roundtrip tests used np.random.default_rng() with no seed, so each CI run drew different data and results were non-reproducible.
2. Too-tight tolerance -- the ideal quantization error is at most 0.5 * scale, but the source original is float32 and the dequantized output is also cast back to float32. Each carries up to ~max_value * eps_f32 of representation error on top of the rounding error. The measured |dequantized - original| therefore occasionally exceeds the exact 0.5 * scale bound (about 0.8% of random seeds for the float64-intermediate case), tripping the assertion with rtol=0.

Fix

• Seed the RNG (np.random.default_rng(0)) in all three quantized roundtrip tests so runs are deterministic and reproducible.
• Widen the tolerance for the float64-intermediate uint16 case to 0.5 * scale + 2 * max_value * eps_f32, reflecting the real float32 representation error budget. The other two cases already used looser bounds (scale / 1.0 * scale).

Verified the widened bound holds across 5000 seeds (worst observed error 0.000500 vs atol 0.000514), so the test stays green even if the seed changes later.

Testing

bazelisk test //ncore/impl/data/v4:pytest_components_3_11
//ncore/impl/data/v4:pytest_components_3_11   PASSED