Summa: fix broadcast-root world rank to include the grid offset

[evaleev]

Problem

Summa::get_row_group_root/get_col_group_root reconstruct the broadcast root's world rank by hand as rank_row*proc_cols + group_root (row; symmetric for col), omitting ProcGrid::rank_offset_. The sparse group factory make_group inserts that root via ProcGrid::map_col/map_row, which do add rank_offset_. On an offset grid (proc_h_ > 1 — the h-grouped 3-D batched SUMMA), rank_offset_ > 0, so the hand-computed world rank is not the value stored in the group: madness::Group::rank() returns -1, tripping MADNESS_ASSERT(group_root >= 0) in WorldGopInterface::bcast → SIGABRT.

The broken branch is only reached when the broadcast group is pruned below proc_cols/proc_rows (a block-sparse operand), so the bug needs both a pruned sparse group and a nonzero offset (batched 3-D grid). Plain 2-D dense grids have rank_offset_ == 0, which is why it stayed latent (and why np=1,2 CI never caught it). Encountered in MPQC PNO-CCSD (block-sparse (g·C) ToT×dense SUMMA) on 8 ranks.

Fix

Compute world_root via the same map_col/map_row primitive make_group uses, so the root is guaranteed to be the in-group owner; TA_ASSERT(group_root >= 0) to fail at the source. Applied to both row and col helpers.

Verification

• Reverting only the fix reproduces MADNESS ASSERTION FAILED at np=8 in general_product_distributed_suite/dist_sparse (passes at np≤2 where proc_h_=1).
• With the fix, dist_sparse and the full general_product_distributed_suite pass at np=4 and np=8; no regressions in dist_eval_contraction_eval/expressions/expressions_sparse at np=2,8.
• New regression: tests/proc_grid.cpp::summa_bcast_root_offset (np=1/2/8) — validates that on an offset ProcGrid, map_col/map_row include rank_offset_ so the root the fix uses is the in-group world rank.