DistArray::lazy_deleter: skip lazy_sync when invoked from fence's do_cleanup

[evaleev]

Summary

Use MADNESS's new WorldGopInterface::is_in_do_cleanup() flag to short-circuit the cross-rank lazy_sync handshake when lazy_deleter is called from inside fence_impl's deferred-cleanup phase: delete pimpl directly, decrement cleanup_counter_, return.

Also bumps the MADNESS pin to m-a-d-n-e-s-s/madness#695 to pick up the flag.

Why

lazy_sync's purpose in lazy_deleter is to keep a peer from sending AM addressed to a WorldObject after the local rank has deleted it. When lazy_deleter runs inside fence_impl's do_cleanup:

• fence_impl's global-termination protocol (drain loop + Dykstra-style nsent/nrecv tree handshake) has already established that no AM is in flight and all ranks are at the same syntactic point.
• defer_deleter_to_next_fence is, by contract, used collectively, so every rank's deferred list holds the same pimpls at this point and every rank reaches the same delete in lockstep.

The handshake is therefore redundant — and actively harmful. The lazy_sync_children task it schedules lives on this world's taskq after the drain loop has exited. That task survived the fence and waited for matching lazy_sync calls that could come arbitrarily later — or not at all, if the world was torn down first. The most visible symptom: ephemeral worlds (e.g. TA-einsum's per-Hadamard sub-Worlds) being torn down at function exit or during exception unwind, leaving the task to be picked up by some unrelated later fence which runs delete pimpl against a freed world. ~WorldObject then trips its World::exists(&world) assertion and aborts.

The general (non-deferred) path is unchanged: lazy_deleter invoked outside do_cleanup still routes through lazy_sync because we cannot rely on synchronization with peers in that case.

What

• external/versions.cmake: bump pinned MADNESS to 666765ca6.
• src/TiledArray/array_impl.h: in lazy_deleter, after the existing world.await(num_live_ds == 0), check world.gop.is_in_do_cleanup(); if set, delete directly.

Test plan

• Compiles cleanly against MADNESS#695.
• Local downstream MPQC repro that previously aborted in ~WorldObject no longer needs any einsum-side workaround to run end-to-end (verified by toggling).
• CI green.

Follow-up

ValeevGroup/tiledarray#550 builds on this with a smaller einsum RAII guard and the (outer Contraction, inner Hadamard) view-cell case in cont_engine.