Add Sequence space, update flatten functions

[Markus28]

We are trying to integrate some of the functionality from https://github.com/entity-neural-network/entity-gym into Gym. We agreed that the best first step would be to implement a Sequence space that models arbitrary-length sequences.

This leads to some problems because we cannot flatten arbitrary-length sequences to a Box in a reasonable way. The same problem exists for the Graph space. Currently, flattening a graph just means flattening the node and edge feature spaces. This seems quite reasonable to me and I took the same approach for the Sequence space. However, some changes were necessary to make flattening work for Dict and Tuple spaces that also contain Graph or Sequence spaces.

TODO

• New tests for flatten functionality, Sequence space
• Improve docstrings

[pseudo-rnd-thoughts]

Looks good, a minor point on the seeding
Would you be able to check this against my updated seed testing PR, this should simplify the testing additions
#2977

Are we adding the utils.py function support? Otherwise, just raise an error explaining why I think

[pseudo-rnd-thoughts]

Normally the seeding for composite spaces includes the super seed list

gym/gym/spaces/dict.py

Line 103 in 8b74413

def seed(self, seed: Optional[Union[dict, int]] = None) -> list:

[Markus28]

Are you saying that I should change the typehint to seed: Any = None? I can see how that makes sense.

[pseudo-rnd-thoughts]

I think this is what I mean

seeds = super().seed(seed)
seeds += self.feature_space.seed(seed)
return seeds

[Markus28]

Ohhh, of course. I'm adding that :)

[pseudo-rnd-thoughts]

LGTM, the number of tests is impressive.
Could we confirm with the entity space team that this is good before merging

[Markus28]

It might actually make sense to move the spaces and samples from test_utils.py to a separate file for better readability

[cswinter]

LGTM!