Support Dict observation spaces in black_death

[Lauqz]

Description

black_death_v3 asserts every agent's observation space is a Box:

observation sapces for black death must be Box spaces, is Dict(...)

It needs that because it substitutes a zero observation for dead agents and
builds it with np.zeros_like(space.low).

That rules out any environment where an agent observes more than one thing at
once, which is common in robotics: a camera image plus a position vector, say.
Those environments cannot use black death at all unless they flatten first,
which defeats having a structured observation in the first place.

This adds Dict support. The check recurses into Dict so a space is accepted
when every leaf is a Box, and the zero observation is built with the same
recursion so it keeps the shape of the space instead of collapsing it. Box is
unchanged and nested Dicts work.

Also fixed the typo in the assertion message, sapces to spaces.

I hit this on a vision-based multi-agent setup (PettingZoo + SuperSuit + SB3 over
AirSim/UE4) where each drone sees an RGB frame and its own 3D position.

Testing

Three tests in test/parallel_env_test.py: reset and step with Dict
observations, an agent missing from the step output getting a correctly shaped
and correctly keyed zero observation instead of an error, and a Dict holding a
MultiBinary still raising.

Full suite against main as a baseline:

	result
main	117 passed, 46 skipped
this branch	120 passed, 46 skipped

Difference is only the three new tests.

I also ran just the black_death and frame_stack tests from
test/pettingzoo_api_test.py, which are parametrized over atari, butterfly,
classic, mpe and sisl. Identical 123 passed, 51 skipped on main and with this
change, so nothing changes for existing environments.