Polish: seeded rollout-buffer RNG and eval-freeze for the normalization wrappers

[DenisDrobyshev]

First batch from a correctness pass over the library. The core is in good shape: I went
through the on-policy and off-policy training loops, the GAE and n-step return
computations, SAC, TD3, and IMPALA's V-trace, plus evaluate_policy and the rliable-style
evaluation statistics (IQM, bootstrap CI, probability of improvement), and found them
correct. The two fixes here are both in the utility layer.

Seeded rollout-buffer RNG

RolloutBuffer shuffled its minibatches with the global NumPy RNG. Two on-policy agents
constructed with different seeds in the same process therefore shared one shuffle stream
and perturbed each other, and the buffer never actually held the per-instance RNG the
README credits every buffer with. It now owns a numpy.random.Generator seeded from the
agent's seed, threaded through the on-policy base and IPPO. PPO is reproducible per seed
and independent of global RNG state; a regression test covers the seeded, isolated, and
reordering cases.

Eval-freeze for the normalization wrappers

NormalizeObservation and NormalizeReward updated their running statistics on every step
with no way to stop, so using them during evaluation kept the statistics drifting toward
the evaluation distribution and reported reward-normalized returns instead of real ones.
Both now take set_training(bool): frozen, the observation wrapper normalizes with the
statistics learned during training and the reward wrapper passes the environment's original
rewards through untouched. This matches the freeze semantics of Stable-Baselines3's
VecNormalize.

Tests, docs, and the CHANGELOG are updated. ruff, mypy, and the affected tests pass locally.