Plane-flow configuration carried by CubeDimResource::Specialized. Holds the
counts for main-flow vs load-only planes and the partition rule used at
runtime.
The instruction shape to contract with, chosen from the problem’s aspect ratio:
32x8x16 for a tall problem, 8x32x16 for a wide one, 16x16x16 for a balanced
one, then 8x8x8, then whatever the device supports at all.
Returns the (row, col) of the current compute primitive within the stage,
deriving compute_index from role_rule_config via PlaneFlowPartition
and delegating the per-scope math to
Partitioner::coordinates.