perf: add radix sort for GSplat tile sorting on NVIDIA GPUs

[mvaligursky]

Add a per-tile radix sort pass as a faster alternative to bitonic sort for tiles with ≤1976 entries on NVIDIA GPUs. Uses a 5-pass, 4-bit radix sort with subgroupBallot-based stable scatter, operating entirely in 16KB of workgroup shared memory.

Changes:

• Add workgroup-local radix sort shader using subgroupBallot with .x-only optimization (correct for sgSize ≤ 32)
• Three-tier tile classification when radix is enabled: radix (≤1976 entries), bitonic (1977–4096), large (>4096 via bucket+chunk)
• Gate radix sort to NVIDIA only — Apple shows a performance regression vs bitonic; AMD wave64 (sgSize 64) would produce incorrect results with .x-only ballot; other vendors untested
• Add GPU vendor detection via capsDefines (VENDOR_NVIDIA, VENDOR_APPLE, etc.) by overriding initCapsDefines() on WebgpuGraphicsDevice
• Rename bitonic sort profiler labels to GSplatLocalTileBitonicSort / GSplatLocalChunkBitonicSort for clarity

Performance:

• On NVIDIA (RTX 2070 Super), radix sort is faster than bitonic for tiles with ~1000+ entries, with increasing advantage at higher splat counts
• On Apple M4 Max, radix sort shows a 5–50% regression vs bitonic — excluded via vendor check
• No performance impact on non-NVIDIA hardware (radix path is completely skipped)