Add: child_memory + TensorKey + scheduler affinity for device-resident tensors

[ChaoWao]

Summary

Enable device-resident tensor buffers to flow through the L3 task pipeline without redundant H2D copies. Adds memory allocation on next-level workers via orch.malloc, scheduler worker affinity, and TensorMap compound keys for cross-NPU address disambiguation.

1. child_memory flag on ContinuousTensor

A 1-byte field in existing tail padding (sizeof stays 40B). When set, init_runtime_impl passes the tensor pointer through as-is instead of device_malloc + copy_to_device + record_tensor_pair.

2. device_malloc_ctx C API

Export device_malloc_ctx / device_free_ctx / copy_to_device_ctx / copy_from_device_ctx from pto_runtime_c_api (all 4 platforms). Wired through ChipWorker.malloc/free/copy_to/copy_from.

3. orch.malloc(worker_id, size) — memory allocation on next-level workers

Full path: Orchestrator → WorkerManager → WorkerThread → ChipWorker.

• THREAD mode: direct call on ChipWorker (MemoryAllocator is mutex-protected for thread safety)
• PROCESS mode: control command via mailbox IPC (CONTROL_REQUEST/CONTROL_DONE states)

4. TensorKey compound key for TensorMap

Replace uint64_t key with {uint64_t ptr, int8_t worker} struct. Disambiguates identical device addresses across different NPUs.

5. Scheduler worker affinity

• submit_next_level(worker=0) / submit_next_level_group(workers=[0,1]) store per-args affinities
• Scheduler::dispatch_ready two-pass: satisfy affinity constraints first, then fill unconstrained slots from idle pool

6. Thread-safe MemoryAllocator

Add std::mutex to MemoryAllocator (all 4 platform implementations) so orch.malloc from the orch thread can run concurrently with init_runtime_impl on the worker thread.

Testing

• C++ unit tests pass (8/8) — child_memory roundtrip + TensorKey compound key + orchestrator
• Python unit tests pass (7/7 new + existing)
• L3 scene test passes on a2a3sim — orch.malloc → orch.copy_to → child_memory=True → submit_next_level(worker=0) × 2
• Hardware tests (requires device)

Related: #571

[gemini-code-assist]

Code Review

This pull request introduces a child_memory flag to the ContinuousTensor structure, allowing the runtime to skip host-to-device copies for tensors already managed by a child process. The changes span the core task interface, Python bindings, and runtime implementations for multiple platforms. Feedback highlights a critical regression where skipping these tensors in the initialization phase causes misalignment in the result validation logic, potentially leading to data corruption. Furthermore, the provided system test uses host-side shared memory to simulate device memory, which may not accurately represent behavior on physical hardware.