feat(env): add NumPy manager-based lifecycle

[TATP-233]

Summary

• add the canonical NumPy ManagerBasedRlEnv / ManagerBasedRlEnvCfg lifecycle on top of the single existing NpEnv.step/reset/autoreset path
• preserve NpEnvState, explicit policy/critic observation mapping, timeout semantics, final observations, manager ordering, recorder hooks, and env-owned RNG/control buffers
• fail closed at construction for unsupported mjwarp multi-substep callbacks and per-substep metrics; do not add legacy DR, backend-private, IPC, runner, or fallback paths

Scope accounting

• 9 files changed: +1,229 / -22
• source-derived copied code: 0 lines (the pinned mjlab lifecycle semantics are implemented through the existing UniLab NpEnv; no second upstream loop is vendored)
• mechanical Torch → NumPy adaptation: 0 standalone lines in this child (completed in the manager-package child)
• UniLab-specific implementation: 718 production additions; adapted/focused tests: 511 additions
• modified existing: 7 files; new: 2 files; deleted: 22 lines / 0 files

Validation

• focused manager/lifecycle tests: 123 passed
• representative regression: 387 passed, 14 skipped, 1 xfailed
• make test-all: 1,833 passed, 24 skipped, 1 xfailed; ruff, mypy, and pyright passed; benchmark smoke 65/67 with only the two optional mlx cases skipped

Implements #1049
Umbrella: #1042