Simplify, fix, and add unit tests for PipProvider.get_preference

[notatallshaw]

Fixes: #12975

This PR does the following:

• Removes direct as a preference based on this reasoning: direct preference in resolution doesn't work #12975 (comment)
• Simplifies the code in get_preference (removes logic flow via exception)
• Adds a unit test for each preference in get_preference
• Small update to the docs for get_preference

Side note: I think I've found a bug in the way parent_depths is calculated when a requirement with an extra is involved, but it was too involved to fix in this PR, once I've pinned it down I will make a follow up PR.

[notatallshaw]

CI is green, ready for review / merge

[notatallshaw]

Tentatively putting this on the 24.3 milestone since it's a small bug fix, a small refactor, and finally unit tests get_preference! Feel free to remove.

[sbidoul]

For this one, we'd need a review by a committer who is more familiar than I am with the innards of the resolver.

[notatallshaw]

I ran all scenarios against: https://github.com/notatallshaw/Pip-Resolution-Scenarios-and-Benchmarks

There is no regression or advantage between this PR and pip 24.2, which is what I expected (there was a small chance removing the broken direct preference would change something, but at at least in all these scenarios it didn't).

[sbidoul]

This sentence seems to be incomplete?

[notatallshaw]

Thanks, I've update to match the style of the surrounding sentances.

[sbidoul]

Is this sentence correct?

[notatallshaw]

Thanks, I've updated this point to match what I wrote in the docs

[notatallshaw]

Going to make a single follow up PR once #13001 lands, I'll comment here once done.