Fix CUDA stream reuse in clue filter

[mikemcqueen]

Summary

• fix CUDA stream reuse so per-stream source counts are refreshed when later sums shrink
• initialize stream stride before device setup and zero stream device pointers
• size filter scratch buffers to active source count instead of full occupancy

Verification

• node-gyp rebuild
• node clues -pf.72 -c2,5 -m3 --xor constitution > /dev/null
• node clues -pf.72 -c2,5 -m3 --xor constitution -v > /dev/null