Improve server BSP performance

[cyberjunk]

This is a cleaned up variant of https://github.com/OpenMeridian/Meridian59/pull/1344 It does not contain any extensive additional benchmark projects etc. It contains the following improvements:

1) https://github.com/OpenMeridian/Meridian59/commit/8dde32766c406851b4a478af080436684d15e7fd This will add an improved geometry.h version (fully backward compatible), containing the following improvements:

• Provides faster explicit SSE implementations for some vector-operations. These can be enabled by the preprocessor macros (SSE2 or SSE4). Fully disabled by default, because I think we should make sure all the other changes are fine first. Sadly the required SSE4.1 instructions are not supported on AMD until FX/A10 (no Phenom II). So a possible default preprocessor macro would likely be "SSE2", but "SSE4" can give nice boosts on Intel hardware.
• The ISINBOX() macro check has been ported from slow calls to "minf()" and "maxf()" to fast macros MIN() and MAX().
• Some new vector functions (such as for rounding a vector) have been added and optimized for SSE4.
• Some of the more complex intersections algorithms in it have been slightly improved.

2) https://github.com/OpenMeridian/Meridian59/commit/074b0ee51eca39d6abc0bca0b724474c2b871def This contains the following improvements to the BSP queries in roofile:

• Replaced recursive 'GetHeightTree()' function with a simple and a lot faster while loop. This is very easy in this case, because you always have to move down to one child only. This should speed up the GetLocationInfo() and the LOS (which retrieves both heights first) a lot.
• Moved some duplicated code to intersect a finite line (los or move) with a BSP splitter into an inlined function (was used in both, LineOfSight and CanMove)
• Uses the new "AllocateMemorySIMD()" variants from memory.c. It is very important to allocate any struct, which directly or indirectly (nested) contains a V2/V3 member by using this function (of course this is not true for pointer members to V2/V3). If SSE2/SSE4 is disabled, the call simply maps to the default allocate call, if it's is enabled, it will use the aligned allocation call. V2 and V3 allocations will both require 16 bytes memory then (instead of 12 and 8).

---

Other things/improvements: 1) Some logical operators (in geometry.h as well as roofile) have been replaced by binary ones, in cases the expression is very unlikely to skip over some further conditions. 2) Linux compatibility: Not tested so far

---

Further things to improve (not in this pull): 1) The "constant" LOS or Move line S->E is calculated on the fly a lot during the recursive calls. Especially for each triangle-intersection right now It might be better to do this once before the recursion/algorithm starts. 2) The IntersectLineTriangle() is suboptimal for the convex polygon check. It has some "intro"-calculation to check whether line and triangle are parallel (which is either true for all or none of the triangles in the convex polygon). Preferrably the geometry.h should contain an "IntersectLineConvexPolygon()" function, which is build from the triangle-intersection algorithm and checks the whole polygon optimized. 3) The recursive tree-traversal functions for LOS and CanMove could be ported to while based ones. Since these functions have cases which require evaluating both children, it's not as trivial as the GetHeight() rewrite.I think there exists a rather complex "constant-space-while-tree-traversal" algorithm, however the readability of these functions could likely be hurt...

[skittles1]

+1, please let me know when you've finished with this PR so we can test/merge

[cyberjunk]

+1, please let me know when you've finished with this PR so we can test/merge

I don't plan to add anything to this pull except bugfixes. Everything mentioned in "Things to improve" would be a next, dedicated pull.

First thing we probably should check now is linux compatibility...