Sample DelayBuffer lags on the first step after creation or reset

[Zhuoxi2000]

DelayBuffer starts every row at lag 0, and reset() sets it back to 0. A row's lag only changes on that row's next scheduled refresh. With update_period > 0 and per_env_phase=True (the default), most rows therefore keep lag 0 for up to update_period - 1 steps. With hold_prob > 0, the first refresh can also keep the 0. As a result, even a constant-delay buffer (min_lag=max_lag=2) returns the undelayed frame for many environments at the start of every episode. This contradicts the documented contract: min_lag is "Minimum lag (inclusive)", lags are "sampled uniformly from [min_lag, max_lag]", and the docs say to use min=max for constant delay. It affects both observation terms and actuators that set delay_*. With the example from docs/source/actuators.rst (min 2, max 5, hold 0.3, period 10, 8 envs, seed 0), 7 of 8 envs have lag 0 for the first 7 steps.

This change adds a per-row flag for rows that have not had a lag sampled or set since creation or reset. Those rows always sample on their next compute(), regardless of update phase or hold_prob, and then follow the normal schedule. A staggered env now samples at t=0 and then at its usual phase. A lag set with set_lags() before the first step is still kept, which test_delayed_actuator_set_lags_affects_delay relies on. reset() still reports lag 0 until the next compute(), so the existing reset tests are unchanged. The change adds no random draws and no host sync.

test_lags_stay_within_range_from_first_step covers the staggered update_period case and the hold_prob case, both at creation and after a full reset. Both cases fail on main and pass with the fix. With the fix, tests/test_delay_buffer.py, test_delayed_actuator.py, test_observation_delay.py, test_actuator_builtin_group.py, test_observation_manager.py and test_actuator.py pass, and ruff, ty and pyright are clean.

#1191 also touches _sample_lags, but only the hold draw for per_env=False. This change merges cleanly with it apart from the changelog entry.

AI assistance: this change was drafted with an AI coding assistant (Claude) and verified locally with the tests above.