arena ToT: einsum sub-World fence guard + (outer-Contraction, inner-Hadamard) view-cell case

[evaleev]

Summary

Stacked on top of #551 (lazy_deleter fast path).

• einsum: add a single-fence RAII FenceSubWorldsOnExit guard so any tasks scheduled by non-deferred ~DistArray calls (e.g. AB during exception unwind) are drained before sub-Worlds are torn down.
• cont_engine: add the (outer Contraction, inner Hadamard) view-cell case for ArenaTensor-backed ToT inner tiles, via the existing arena fast path.

Important

The PR base is set to evaleev/feature/lazy-deleter-skip-sync-in-do-cleanup (#551). Once #551 merges, this PR will auto-retarget master.

Why

Triaged from an abort while running a CSV-CCk traced expression over ArenaTensor ToT operands in MPQC. Cascade:

1. The expression C(i_3,i_4;a<...>) = A(i_3;a<...>) * B(i_4;a<...>) -- the typical sub-product inside einsum's generalized-contraction loop after ToT operands are reduced per Hadamard tile -- hit the view-cell branch of init_inner_tile_op with (outer Contraction, inner Hadamard). That combo had no handler and threw \"nested non-contraction product on view inner tiles is not yet supported\".
2. The exception propagated out of einsum's per-h iteration. The temporary worlds vector destructed during stack unwind, tearing down sub-Worlds while lazy_sync_children tasks scheduled by ~DistArray's lazy_deleter were still in the global ThreadPool's queue.
3. A later fence in the enclosing scope picked up those stranded tasks; ~WorldObject then asserted World::exists(&world) and aborted, masking the real TA exception.

#551 eliminates the deferred-cleanup half of (2) at the source (no lazy_sync task ever scheduled in that path). This PR adds the missing view-cell case so (1) does not throw in the first place, and keeps a thin RAII guard so the non-deferred path during exception unwind also leaves the sub-Worlds clean.

What

einsum RAII guard

FenceSubWorldsOnExit declared right after the worlds vector. LIFO destruction means AB/C destruct first (releasing pimpls; the non-deferred path goes through lazy_sync and enqueues tasks on sub-Worlds), then the guard runs (drains those tasks via a single fence per sub-World), then worlds destructs (empty taskqs). A single fence per sub-World is sufficient because #551 makes the deferred-cleanup ~DistArray path skip lazy_sync (no post-do_cleanup task to drain); only tasks from non-deferred destructors remain, and a single drain suffices.

All participating ranks of a sub-World reach this RAII guard at the same point in lockstep at function exit, so their lazy_sync handshakes match up.

cont_engine view-cell case

init_inner_tile_op now handles (outer Contraction, inner Hadamard) for view inner cells. Mirrors init_inner_tile_op_owning_: arena_plan_ with ArenaInnerShapeKind::left_range and a per-cell make_fused_hadamard_lambda / make_fused_hadamard_scaled_lambda op that accumulates r += l * rr (optionally scaled) into pre-shaped view cells. Non-identity inner result permutation is rejected explicitly (the owning fallback that materializes a permuted return cell cannot run for views).

Test plan

• Local downstream MPQC repro (bimal.json CSV-CCk trace evaluator): previously aborted in ~WorldObject during stack unwind; now runs end-to-end with the real TA expression evaluating cleanly, exit 0.
• CI green (single-rank and multi-rank).