…atch==1 assert) + tile tests

…ntEngine + e2e (nbatch>1)

…) + tile tests

…ContEngine + e2e (left-external Mo>1)

…left-external + nbatch>1) + benches

Route the regime-A hc+e einsum (outer Hadamard + outer contraction,
inner outer-product) through the landed arena_strided_dgemm_ce_e core
(M=N=1, K=tile volume) in run_regime_a_arena, replacing the per-cell
rank-1 dger loop with one strided DGEMM per outer-contraction tile.
Gated to view+double arena ToT contraction with num_contract_ranks()==0;
all other kinds keep the per-cell path. Adds a regime_a_strided_disabled()
kill switch, tile/e2e/differential/edge tests, and a strided-vs-per-cell
benchmark (~7.3x on a C6H14-like shape).

… (either-side hce+ce)

… timing probe, bench & tests

… ranks

The einsum_tot arena-matches-owning tests iterate over all result tile
ordinals but only inspect tiles local to the calling rank, then assert
the per-rank elements_compared / result_outer_cells_seen counts (and the
fatal BOOST_REQUIRE_GT(elements_compared, 0u)) against the global
expected totals. That holds under np=1 (all tiles local) but fails under
np=2: each rank sees only its share, and a rank owning no result tiles
trips the REQUIRE_GT.

All-reduce the accumulators (gop.sum on the counts, gop.max on
max_abs_diff) before the assertions so every rank checks the true global
totals. Fixes the 14 np=2 einsum_tot failures.