Add intelligent fallback from backjump to backtrack

[notatallshaw]

Fixes #134

Alternative to #142, fixes both the requirments mentioned here #142 (comment)

I beleive this strategy is fairly foolproof:

1. When backjumping make a backup of the states the first time backjumping skips over states further than backtracking would
2. If backjumping reaches resolution impossible and a backup state exists, fall back to that state and start backtracking

However, there are downsides, it could definetly cause a valid ResolutionImpossible to take much longer to produce. I have therefore give the calling library a choice on what to do, backjump with backtrack fallback, backjump only, or backtrack only.

If you want tests I need some small guidance, I don't really understand this repos tests.

Let me know what you think!

[notatallshaw]

FYI a seperate Pip PR (pypa/pip#12459) I have removes all examples I have tested for backjumping incorrectly producing ResolutionImpossible, I think it's important to still remove or have a fallback for backjumping, I have a more detailed comment in the alternative PR #142 (comment)

[frostming]

The implementation is a bit convoluted, need more comments to help successors understand what is happening.

[notatallshaw]

The implementation is a bit convoluted, need more comments to help successors understand what is happening.

I think to a certain extent that is the nature of this solution, it backs up states and only uses those backed up states under the right conditions.

But I have tried to improve the situation breaking up the code a little, adding more comments, and trying to pick better names. Let me know what you think.

[notatallshaw]

An further thoughts on this?

It would be good to fix this backjump issue somehow, as it breaks real world requirements for both pip and PDM.

[notatallshaw]

I am unhappy with this solution, I am working further on understanding the root cause of the problem: #134 (comment)

I may recreate this PR once I have a working test.