Make KD-tree decoder memory linear in the point dimension

[Filyus]

DynamicIntegerPointsKdTreeDecoder allocates memory quadratic in the point dimension, up front, in its constructor. The dimension is the total number of components over all attributes, and a stream sets it at five bytes per attribute. This change makes the decoder's memory linear in the dimension. The decoded output is unchanged.

Decoder memory by dimension:

dimension	main, always allocated	this change, at most
3	2.3 KB	1.6 KB
64	1 MB	33 KB
255	17 MB	133 KB
2,048	1.1 GB	1.1 MB
6,375	10 GB	3.3 MB

What this means for actual streams:

• A 143-byte point cloud with 25 attributes of 255 components makes main commit about 10 GB before it reads any KD-tree data, and then fail on the missing data. With this change it fails without that allocation.
• A valid 8,372-byte stream of three points with 2,048 components decodes with about 5 MB in 80 ms, against 1.1 GB and 1 to 1.9 s on main.
• Ordinary point clouds decode as fast or faster. Best of 200 decodes, Release build with MSVC:

level	components	points	main	this change
0	3	50,000	1,722 us	1,705 us
2	3	50,000	1,967 us	1,921 us
6	3	50,000	2,032 us	1,894 us
4	6	20,000	1,007 us	948 us
2	64	2,000	487 us	338 us

How: the decoder kept a full copy of the node's base and levels for every level of the tree, which is 32 × dimension levels deep. A split changes only one base value and one level. The decoder now keeps a single base and levels vector, records the old values at each split, and restores them when the walk returns to a shallower node.

Verification:

• The new test fails on main (AddressSanitizer runs out of memory in the constructor) and passes with this change. draco_tests passes under AddressSanitizer.
• 1,200 generated point clouds over all compression levels and 1 to 100 components, each decoded as encoded and with 30 damaged copies, give the same result and the same attribute bytes on main and on this change. That probe is not part of the change.
• Nine existing PointCloudKdTreeEncodingTest cases fail if the restore is broken, so the existing tests cover it.