refactor!: use ops._main._Manager in Scenario

[tonyandrewmeyer]

The ops._main._Manager class started out as the Ops class in Scenario, but went through significant modifications before it was added in ops. This PR merges those changes back into Scenario, so that it's also using the _Manager class. Other than generally simplifying the code, this also means that Scenario is even more consistent with using ops in production.

This includes some small refactoring in ops to make this cleaner:

• ops._main._emit_charm_event is only used by _Manager, so logically belongs inside that class. This also makes it much simpler to have a subclass override the behaviour.
• Allow passing a JujuContext to _Manager, so that it can be populated by something other than os.environ.
• Allow _Manager subclasses to override how the charm metadata is loaded.
• Allow _Manager subclasses to execute code after committing but before the framework is closed.
• Support a dispatch path that uses _ rather than - in event names (Scenario consistently uses _, even though this would always be - from Juju - feel free to push back on this, requiring Scenario to undo the _ conversion when populating the context).

The Scenario runtime and ops_main_mock modules are renamed to properly indicate that everything inside of them is an implementation detail and should be considered private by Scenario users. This is a breaking change but we're considering it a bug fix and not one that requires a major release, given that the documentation and top-level namespace were already clear on this.

Significant Scenario refactoring:

• UnitStateDB is moved from runtime to ops_main_mock, and it no longer (re) opens the SQLiteStorage - it instead takes a reference to the one that's in the framework, so that we're consistently using the same underlying storage, without relying on the filename being consistent (this eases moving to :memory: in the future).
• As far as I can tell, runtime._OpsMainContext is completely unused and missed being removed in previous refactoring, so it's entirely deleted.
• Loading and storing deferred events and stored state into the state database is moved from the runtime to the manager subclass.
• The Ops class survives, but is now a subclass of the ops._main._Manager class. A reasonable amount of functionality is now removed (because the parent handles it), and the remaining code is specifically for handling the testing framework.

A few type hints are added in some tests where I was working on fixing the test and it helped to do that.

Fixes #1425

[tonyandrewmeyer]

This maybe broke the recent variadic type change, since ops._main._Manager now sets the .charm attribute and the change didn't extend that far. I'll look into that.

[james-garner-canonical]

Looks like a bunch of changes have come in while I'm reviewing, I'll just leave these comments for now and finish the review later.

[james-garner-canonical]

If snake case is only being produced by scenario, it feels kind of odd having this one special case in ops._main.

Changing scenario to produce kebab-case instead of snake_case seems like it could be a pain though -- and would it be a breaking a change for users? So probably not worth doing that.

Would it be nice to have a standard way of handling snake vs kebab?

if kebabify(self._juju_context.dispatch_path).endswith('-relation-broken'):

I don't really love this either.

I don't think this topic should really be a blocker either way.

[tonyandrewmeyer]

Looks like a bunch of changes have come in while I'm reviewing, I'll just leave these comments for now and finish the review later.

The changes are only (meant to be!) merging in the charm type PR (well, rebasing to get that, but because I rename the file here it was a bit messy). No functional changes, although as noted above I wonder if I break that feature here and need to get the generic type over to main somehow.