Fix: separate onboard run context cleanup from final device teardown

[ChaoZheng109]

Summary

• Remove resource cleanup from run_runtime() lambda — let finalize_device() handle it via DeviceRunner::finalize()
• Add rtSetDevice() at finalize() entry so rtFree succeeds even when called from a thread without an existing device context
• Move rtDeviceReset() from reset_device_context() into finalize(), after mem_alloc_.finalize(), to guarantee all device memory is freed before the context is torn down
• Drop MemoryAllocator::finalized_ flag — an empty ptr_set_ is already a natural no-op, so repeated and interleaved finalize() calls are idempotent without an explicit guard

Root Cause

The original teardown sequence was:

1. run_runtime() lambda → reset_device_context() → destroys streams + rtDeviceReset
2. Outer finalize_device() → DeviceRunner::finalize() → mem_alloc_.finalize() → rtFree

Step 2's rtFree calls happened after rtDeviceReset in step 1, causing error 507899 (invalid device context).

The fix centralizes all resource cleanup in finalize(), called by finalize_device():

finalize_device() (main thread):
  → finalize()
    rtSetDevice(device_id_)    ← establish context
    streams + rtFree           ← context valid
    rtDeviceReset(device_id_)  ← last step

Testing

• All 19 a2a3sim examples pass

[gemini-code-assist]

Code Review

This pull request refactors the device cleanup logic by moving the rtDeviceReset call from reset_device_context to the finalize method in DeviceRunner. It also updates the MemoryAllocator to reset its finalized_ state upon new allocations and modifies the runtime API to call finalize instead of reset_device_context. Feedback was provided to guard the rtDeviceReset call with a check for a valid device_id_ to ensure the finalize method is idempotent and avoids passing an invalid ID to the runtime.