perf: vectorize Gaussian evaluation across 2x2 pixel quad in rasterize pass

[mvaligursky]

Vectorize the per-splat Gaussian evaluation in the tile rasterizer's color path.

Changes:

• Remove evalSplat function and inline the evaluation directly in the batch loop
• Compute dx = p00 - center once, then build vec4f pixel offsets exploiting the regular +1 grid pattern of the 2x2 quad
• Evaluate power, gauss, alpha, and transmittance update as vec4 operations instead of four independent scalar chains
• Pack transmittance into a single vec4<half> T (was four separate half variables) for branchless alpha-blend update via select() and the saturation early-out check
• DEPTH_TEST mask folds into the vectorized valid condition
• Pick mode (evalSplatPick) unchanged — its complex branching doesn't benefit from vectorization

Performance:

• Neutral on Apple M4 (2.95ms) where GPU ALUs are scalar
• Expected improvement on NVIDIA/AMD where wider SIMD lanes can exploit the vec4 operations
• Eliminates redundant center subtractions (was computed 4× per splat, now 1×)
• Cleaner code: fewer variables, select() instead of manual cond * a + (1-cond) * b