fix: Test manifest to respect parameterized tests

[PragTob]

Fixes #15820

In short, parameterized tests weren't accounted for in the test manifest, so if a variant first failed it could get overridden by a later run with different parameters as the key was just {module, test}.
So, this PR introduces the {module, test, parameters} variant, to account for it.

The old {module, test} is still supported instead of making it {module, test, %{}} as it is part of the official documentation (:only_test_ids) and so we don't want to break compatibility.

Decisions

Some may be worth revisiting:

• bumped the manifest version as we changed it
• decided to support both {module, test} and {module, test, parameters} (see above) but we could also transform it at the boundary (i.e. transform {module, test} input to {module, test, %{}}) and then have a more uniform manifest (and less branching code paths)

Potential Problems

• I believe right now if a parameter gets deleted from a test, a corresponding failure may not be pruned from the manifest as we don't get that data (it's not part of the loaded test module), so then mix test --failed would run but not find any test. Not sure how to best fix that.

[josevalim]

Thank you @PragTob!

I think there is another solution here which is to, in case of parameterized tests, we only write it as passing if all parameters succeed. This way we don't need to change the test ids representation and we always run all parameters for the failing test (which is quite reasonable). WDYT? Perhaps worth trying in another PR?

[PragTob]

@josevalim 👋

Of course, that's a very wonderful idea which would avoid us having to change the manifest! I like the simplicity, thank you!

I'll wip that up later today/tomorrow and we can check then :)