Optimize some parts of NavMap::sync

[Zylann]

This is a few optimizations I've done while investigating navigation on a large procedural streaming terrain.

• Reduced the amount of HashMap lookups
• Use a "pair" structure for connections instead of heap-allocated Vector<T> because they only ever have up to 2 elements
• Pre-allocate instead of letting several reallocations occur in push_back and HashMap re-hashs
• Use squared distance instead of distance to avoid square root calls

Note: I saw that in Find the compatible near edges there is a double for loop, which seems to pretty much iterate every possible pair of free edges. This will iterate each pair twice, (A, B) and (B, A). Unless I'm missing something, the inner loop could use for (unsigned int j = i + 1; j < free_edges.size(); j++) { so each pair only has to be iterated once. But I haven't done that change because I'm not sure if the current loop is actually intented.

[smix8]

Approve for the change in general. This still needs someone code review and a rebase / snyc with #93005 that edits similar parts.

Nit imo would be good to change the unsigned int to uint32_t as that is used for all other LocalVector loops so things stay consistent.

I had time to test this PR more thoroughly with current master and because it reduced the amount of allocation it is a small net performance gain without complicating the code base too much. Might want to move things later to a struct to make the HashMaps and LocalVectors also reusable to keep the memory available but that can be done with another PR. I didnt have time to really look into the link loop but that is definitely an area that also needs improvements at some point as it is a major bottleneck that does not scale well.

[smix8]

@Zylann PR #93005 has now been merged. Would you be available to rebase on top?

[Repiteo]

Thanks!