perf: remove workgroup-level early-out atomics from rasterize pass

[mvaligursky]

Remove the per-batch workgroup-level early-out (atomicStore/atomicAdd/atomicLoad + barriers) from the tile rasterizer. WGSL lacks a fused barrier+vote intrinsic like CUDA's __syncthreads_count, so emulating it with atomics costs 3 synchronization points per batch of 64 splats.

Changes:

• Remove doneCount atomic and doneCountShared workgroup variables
• Remove the per-batch atomic early-out dance (atomicStore → barrier → atomicAdd → barrier → atomicLoad → workgroupUniformLoad)
• Keep per-thread threadDone flag that skips evaluation via branchless ALU once all 4 pixels saturate
• Simplify batch loop to: load → barrier → eval → barrier
• Remove unused WORKGROUP_SIZE constant

Performance:

• Neutral on Apple M4 (2.95ms → 2.99ms) where barriers are nearly free
• Expected improvement on NVIDIA/discrete GPUs where each barrier stall is significant
• Simpler code, less shared memory usage (no atomic variables)