8c969ce7dc
Three changes from reading CCCL dispatch_reduce.cuh + kernel_reduce.cuh +
agent_reduce.cuh + grid_even_share.cuh + summary_statistics.cu:
1. V2 dispatch restored (was hardcoded use_v1=True)
CCCL two-path: single-tile vs multi-tile (GridEvenShare).
Threshold now uses BI-V100 SM count (16) for saturation calc.
2. _forward_decode_pytorch rewritten with CCCL patterns:
agent_reduce ConsumeFullTile: reduced .contiguous() from 4 to 2.
GridEvenShare RAKE tiling: adaptive _MAX_TILE_BLOCKS=1024.
summary_statistics.cu compound reduce: online softmax {m,l,o}.
3. KV gather: permute(1,2,4,0,3) for K avoids intermediate alloc.
CCCL files read: dispatch_reduce.cuh, kernel_reduce.cuh,
agent_reduce.cuh, grid_even_share.cuh, summary_statistics.cu,
kernel_scan.cuh