remove generics in experimental module

[d-v-b]

this PR adds a new top-level module experimental that contain versions of GroupSpec and ArraySpec with the following changes:

• no more generic type parameters
• GroupSpec.members is not nullable. Instead, GroupSpec inherits from BaseGroupSpec, which doesn't have a members attribute. This means we can statically distinguish between a model of a model of a zarr group in isolation (BaseGroupSpec) from a model of a Zarr group as a node in a hierarchy (`GroupSpec).
• allow declaring the full type of the keys + values of members of a GroupSpec. This means enabling the use of typeddicts with specific keys to declare that certain members must have specific names.

I think this design is an improvement over the current ArraySpec / GroupSpec models, but I also want to improve gracefully, so my current thinking is to iterate on these new models in the experimental module and then at some point in time make the breaking change of replacing the models in the top-level .v2 and .v3 modules with the stuff formerly defined in experimental.

curious to get thoughts on that plan / anything else @dstansby @tlambert03

[d-v-b]

@dstansby is it OK if this gets merged + I cut a release? I need the changes in this PR, and the changes are non-breaking

[dstansby]

Sounds good to me 👍 #112 is breaking though, so I think it'll have to be a breaking release regardless of this PR.

[tlambert03]

just looked at the new experimental modules and they look good by me! I think it's a nice simplification: less confusing, and the lost typing subtleties can be worked around in more explicit ways. Would also like to see a new release :)

[tlambert03]

see also #125 (which be merged before a new release)