Fix: store_to_host flush ordering and validation in Worker bootstrap loop

[ChaoWao]

Summary

Follow-up fix for three defects introduced by PR #307 in _chip_process_loop_with_bootstrap's store_to_host DMA wiring. All three stay harmless while buffers are tiny (the allreduce_distributed example currently moves 1 KB per rank), but surface on any production-sized output.

• Race: TASK_DONE was published before the D2H flush ran, so a parent returning from worker.run() could observe the state transition while the chip child still had the flush in flight — zeroed or stale output SharedMemory on real workloads. Flush now runs above _write_error + _mailbox_store_i32(_TASK_DONE).
• Error path: the flush ran unconditionally, including after a kernel failure, stamping undefined device memory into the parent's SharedMemory and masking the real error. Flush is now gated by if code == 0:, and flush-path exceptions convert to code=1 with a store_to_host=<name> prefix instead of propagating and hanging the mailbox.
• Validation: a ChipBufferSpec(store_to_host=True) without a matching HostBufferStaging in host_outputs used to KeyError-crash the chip child after bootstrap reported SUCCESS, leaving the parent polling an uninhabited mailbox forever. ChipWorker.bootstrap_context() now does a symmetric up-front check for both load_from_host ↔ host_inputs and store_to_host ↔ host_outputs so a misconfiguration surfaces as a ValueError on the channel before any device or communicator state is touched.

Test plan

• Added TestBootstrapContextStoreToHost.test_store_to_host_round_trip — 2-rank sim fork that writes a known payload into the window via copy_to, runs the same D2H flush the worker loop does, and asserts the parent's SharedMemory round-tripped the bytes.
• Added TestBootstrapContextMissingOutputStaging.test_store_to_host_without_host_outputs_raises — trips the new symmetry check, asserts child-side ValueError and the parent-visible ERROR channel payload.
• pytest tests/ut/py/test_worker -v (sim)
• python examples/workers/l3/allreduce_distributed/main.py -d 0-1 on hardware (the current — and, until this PR, only — store_to_host consumer)

[gemini-code-assist]

Code Review

This pull request introduces up-front validation for host-staging symmetry in the bootstrap context and refactors the worker process loop to flush 'store_to_host' buffers before publishing the task completion status. This change ensures that the parent process does not read incomplete data from shared memory. The review feedback identifies a potential crash when handling zero-sized buffers in the flush loop and suggests maintaining consistency by using the internal implementation for DMA operations. Similar robustness improvements are recommended for the newly added unit tests.