[metal] Performance Improvements for bitmasked

[k-ye]

• Currently, Metal's listgen will launch a group of threads whose size is equal to that of the total numbers of SNodes. This could be very inefficient. Try grid-stride loops to balance the load.
• Also adopt the grid-stride loop pattern to implement struct_for kernels.
• If the immediate parent of a leaf place is a bitmasked, instead of appending the active bitmasked elements into ListManager, we might want to just loop through all the possible coordinates and check which are active directly (This is how LLVM backends implement it?). This is because the appending to ListManager is expensive, as it uses atomic operations.

[yuanming-hu]

• This is how LLVM backends implement it?

Right. In LLVM ListManager only generates 8x8x8 leaf blocks (instead of every 1x1x1 leaf voxels). This allows the list generation overhead to be amortized by 8x8x8=512 voxels.