Fix multi-agent record_transition to check all agents for episode completion

[darshmenon]

Fixes #289.

Problem

MultiAgent.record_transition uses next(iter(terminated.values())) and next(iter(truncated.values())) to determine which environments finished an episode. This only checks the first agent in the dictionary, so if any other agent terminates first, the cumulative rewards and timesteps are not reset and tracking is wrong.

Fix

Replace the single-agent check with a logical OR across all agents:

PyTorch backend (multi_agents/torch/base.py):

# Before (only first agent checked):
finished_episodes = (next(iter(terminated.values())) + next(iter(truncated.values()))).nonzero(...)

# After (all agents checked via OR):
finished_episodes = (
    torch.stack(list(terminated.values())).any(dim=0)
    | torch.stack(list(truncated.values())).any(dim=0)
).nonzero(as_tuple=True)[0]

JAX backend (multi_agents/jax/base.py):

# Before:
_terminated = next(iter(terminated.values()))
_truncated = next(iter(truncated.values()))

# After:
_terminated = np.stack([jax.device_get(v) for v in terminated.values()]).any(axis=0)
_truncated = np.stack([jax.device_get(v) for v in truncated.values()]).any(axis=0)

An episode is considered finished when any agent signals termination or truncation — consistent with cooperative multi-agent environments where agents share a joint episode boundary.