Add support for python 3.6

[pseudo-rnd-thoughts]

Previously python 3.6 was removed as a supported gym python version, this was primarily as the python foundation no longer supports python 3.6 with security or bug fixes.
This PR readds python 3.6 as a supported version through downgrading importlib_metadata that is used in the registration.py for importing third-party environments.
Looking at the release notes, there appears no good reason for why it was pinned to <=4.10 in eb6d826.
Therefore, we have reduced the strictness to the minimum version that python 3.6 can download, importlib_metadata>=4.8

[RedTachyon]

I mean even if we can support 3.6 now with this simple change, should we? We need to be sure that any dependencies support it too (and if a new release e.g. of pygame with amazing features comes out, but doesn't support 3.6, we need to stick with an old one)

Also, this forbids us from using any 3.7 features. At a glance, this will already break some of the type annotation stuff (PEP 563 which we use in some files), we block ourselves from using dataclasses (or forces us to use weird workarounds; worth noting we already use dataclasses in registration), we can't "safely" use the property that dictionaries are ordered (it's only official since 3.7)

[RedTachyon]

Do we want to specify that it has to be 3.6.15? I'm not sure what the reason for this is exactly, but I'm guessing this should apply to all 3.6.x versions

[pseudo-rnd-thoughts]

This is fine because the python foundation is not releasing any more python versions and this dockerfile is only called by the CI which uses the latest version of each python, i.e. only python 3.6.15 will be used

[pzhokhov]

@pseudo-rnd-thoughts
I am with @RedTachyon on this one -- while we could support python 3.6 (and likely even earlier versions, like 3.5), I'd rather not do it without sufficiently good rationale, for all the reasons that @RedTachyon described, plus generally increased maintenance / development overhead, e.g. anyone developing or modifying a new feature needs to keep in mind that code needs to be python 3.6 compatible. What are the reasons for keeping the python 3.6 support?

[pseudo-rnd-thoughts]

Hi @pzhokhov, thanks for asking.
One of the major users of gym requested that we support Python 3.6 and we would prefer them to utilise the latest versions of gym that includes several important new features than not. Therefore, we are trying to support python 3.6 as a minimal product, i.e. the core API is tested however mujoco is not.
If you join the discord and message me, I can give you more details. I realise that this PR does not explain why we made this choice as the discussions happened privately, hope that helps.

[jkterry1]

@pzhokhov we also warn users to not use 3.6 on import if they're using it, and like Mark said only a very small amount of the library will actually run with Python 3.6

[vwxyzjn]

Hey, @pseudo-rnd-thoughts there is one issue though with python 3.6 support. I am working on the poetry integration as discussed and note that ale-py at minimum supports 3.7 - one of the things to be found out during dependencies resolution :)

Same thing with mujoco https://pypi.org/project/mujoco/#files

This also means if the current user installs gym with python 3.6, it will work but will break if they try to install ale-py.