Improve path finding and radius jewel performance

[vaisest]

Description of the problem being solved:

This PR implements BuildAllDependsAndPaths() and other improvements on top of the previous #10228 PR. In my totally realistic synthetic benchmark this result in around a 3x improvement in its performance. This was originally done to make radius jewel comparisons a bit faster, but they are still very slow, so the difference isn't very noticeable there. I think this is more noticeable on the passive tree where this causes less interface lag.

Note that the performance was only tested with one build, which had a light of meaning, lethal pride, and an impossible escape. Still, it definitely shouldn't reduce performance for any build.

The changes include:

• Caching mastery options (i.e. unallocated mastery info)
• Using multi-source bfs for closest node calculation
• Looking up impossible escape keystone instead of scanning every node
• Avoiding scanning every jewel for each node in spec
• Generally avoiding useless allocations and other hot path inefficiencies similar to the previous PR

Steps taken to verify a working solution:

• Tests pass
• Manually tested features like ascendant pathing which were described in existing comments and code

Link to a build that showcases this PR:

https://poe.ninja/poe1/pob/983a0

[mcagnion]

A local Codex-assisted analysis found a mastery-option cache regression at cf630666. After removing a Runegraft from an unallocated mastery, RemoveTattooFromNode() restores the base node with an empty sd but leaves allMasteryOptions true. BuildAllDependsAndPaths() then skips rebuilding the mastery options, leaving the hover list empty.

A differential test restores 6 option lines on origin/dev and 0 on this PR. A possible fix is to invalidate allMasteryOptions when removing the Runegraft; the same scenario could also be used as a regression test.