Gigaflow reset: agents keep init positions between resamples; resample forces sync resets

[eugenevinitsky]

Splitting out of the discussion on #537 (env.resample_frequency, #537 (comment)) — two coupled issues in the gigaflow reset path:

1. Positions are only sampled at init. An episode-boundary c_reset respawns agents but does not resample scenario positions; a fresh position draw only happens during periodic resampling. So the frequent-resample setting is doing double duty as a position-diversity knob.
2. Resampling forces synchronized resets, which is what makes the logs readable. In training each scene resets when ~40% of its agents are dead (termination_mode 1), so scenes drift async within a rollout and the logs get noisy; the periodic resample forcibly re-syncs every scene. Log cleanliness shouldn't depend on a scenario-resampling side effect.

Desired end state: position resampling happens on every episode reset (decoupled from resample_frequency), and log aggregation is robust to async scene resets so resampling frequency can be chosen for data-diversity reasons alone.

Context: mimolette (and now the defaults from #537) use resample_frequency: 256000 partly because of these side effects.