feat(randomization): add mean_matched_uniform distribution kind

[tkevinbest]

Strictly opt-in: no shipped preset uses the new kind, so every existing config draws bit-identical values.

Two mass randomization presets use plain uniform ranges. link_mass_range: [0.9, 1.2] is a multiplicative scale on each randomized link, unperturbed at 1.0. added_mass_range: [-1.0, 3.0] is an additive kg offset on the torso, unperturbed at 0.0. A uniform's mean is the midpoint of its band, and neither band is centered on its unperturbed value: [0.9, 1.2] averages 1.05, [-1.0, 3.0] averages +1.0 kg. Averaged over environments, the policy trains against a robot whose links are 5% heavier than the URDF describes and whose torso carries an extra kilogram it does not have. A config diff shows none of this; the ranges look reasonable. The asymmetry is deliberate — the wide upper side covers carrying a payload — so narrowing or re-centering the bounds is not the fix.

mean_matched_uniform keeps the same bounds and puts the mean where it belongs. For low < mean < high it draws U[low, mean] with probability (high - mean) / (high - low) and U[mean, high] otherwise, so E[X] = mean exactly while the support stays [low, high].

"added_mass_range": {"kind": "mean_matched_uniform", "low": -1.0, "high": 3.0, "mean": 0.0},

No existing kind does this on an asymmetric band: on [0.9, 1.2], uniform gives 1.05, log_uniform 1.0428, and gaussian with mean=1.0, std=0.05 gives 1.0028, because that parameter is the pre-truncation mean. DistributionSpec.expectation() returns the closed-form mean of any spec, so a range's mean can be checked rather than assumed.

What holds the default behavior fixed:

• Nothing outside distribution.py, sampler.py and their tests references the new kind.
• A bare [lo, hi] pair still parses as uniform.
• The uniform, log_uniform and gaussian branches of _inverse_cdf are untouched. The only edit to shared code moves _INV_SQRT2 and _P_EPS from sampler.py into distribution.py at identical values, so expectation() can share them.
• The sampler is keyed, so switching a preset to the new kind would change drawn values at a fixed seed. That is a separate decision for consumers.

mean is required rather than defaulted to the midpoint, and validation requires strict low < mean < high. A mean at the midpoint routes to the plain-uniform path so the two agree bit-for-bit. The kind string is serialized into checkpoints and DistributionSpec.parse raises on unknown kinds, so renaming it later breaks eval of checkpoints trained with it.

98 unit tests (28 new), no_sim CI selection 551 passed / 9 skipped, pre-commit clean, IsaacSim DR matrix passing. IsaacGym and MuJoCo were not run and the kind has no cross-backend _dr_matrix.py cases, both because _inverse_cdf is the repo's only spec.kind dispatch.