[Feature] Tensor dump for runtime debugging and validation

[ChaoZheng109]

Summary

Add a tensor dump capability that captures intermediate tensor data (inputs before dispatch and outputs after completion) during runtime execution. This enables offline debugging, golden-value validation, and kernel correctness verification without modifying user kernels.

The feature spans three layers:

• Platform layer: common tensor dump interface, AICPU-side dump logic, and host-side collector for gathering dumped data
• Runtime layer: integration into host_build_graph runtime (with future support for aicpu_build_graph and tensormap_and_ringbuffer)
• User interface: --dump-tensor CLI flag in run_example.py and a dedicated example (dump_tensor_example) demonstrating usage

Motivation / Use Case

When debugging kernel correctness issues or validating new orchestration flows, developers currently have no built-in way to inspect intermediate tensor values at each execution step. They must manually instrument kernel code or add ad-hoc print statements, which is error-prone and non-reproducible.

A first-class tensor dump feature allows:

• Capturing before-dispatch inputs and after-completion outputs per task, saved to disk as binary files
• Comparing dumped tensors against golden computations to pinpoint which kernel or step produces incorrect results
• Debugging without modifying kernel source — the dump is controlled entirely from the runtime/platform layer

Proposed API / Behavior

• Enable via --dump-tensor flag on run_example.py
• Runtime sets enable_dump in kernel args; AICPU reads this flag and writes tensor data to a host-visible region
• Host-side TensorDumpCollector gathers and writes binary dump files organized by task ID and tensor index
• Output directory: outputs/tensor_dump_<timestamp>/

Additional Context

Work in progress — currently implemented for host_build_graph runtime on the a2a3 architecture.