Add multi-platform hardware benchmarks

[baominghelly]

Summary

• integrate the final code from hardware_adapt and cambricon_test
• preserve the existing CUDA adapter API, CLI shape, parsers, and metric names
• add MetaX, Iluvatar CoreX, Hygon DCU, Moore Threads, Cambricon, and Ascend support
• keep hardware.cudaUnified as the single hardware framework; select the accelerator with config.device
• keep hardware platform aliases and benchmark configuration in infinimetrics/hardware/constants.py
• exclude internal implementation plans from the release tree
• accept both CUDA and native-platform spacing in memory-bandwidth table headers

This replaces #42 so the published branch follows the repository naming convention. The published single-commit tree is identical to the final seven-platform tested tree:

07db0e9356c422def966aad5743dbec75c530443

Design

Dispatcher continues to route by test category and framework only:

hardware.cudaUnified.<test type>

The hardware adapter resolves the platform from explicit config.device, then from the installed toolchain when the device is omitted. Device names such as metax, iluvatar, moore, cambricon, and ascend are not Dispatcher frameworks.

Validation

• Black 23.9.1: all 78 Python files pass
• A100/Linux: 36 passed
• Python compile checks and git diff --check pass
• all seven platforms ran from Dispatcher-created hardware.cudaUnified adapters with test_type=Comprehensive
• every run used --all --device 0 --iterations 2 --array-size 1048576
• A100: four memory directions, four STREAM operations, L1, and L2 parsed successfully
• MetaX C550, Iluvatar TG-V200, Moore Threads S5000, Cambricon MLU590, and Hygon Z100L: benchmark, adapter validation, and container exit 0
• Ascend 910B4: four memory directions, four STREAM operations, and D2D hierarchy parsed successfully; adapter exit 0, followed by the development container's known cleanup exit 137
• final adapter metrics: 10 on CUDA-compatible/Cambricon platforms and 9 on Ascend