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 verified MetaX, Iluvatar CoreX, Hygon DCU, Moore Threads, Cambricon, and Ascend support
• add multi-platform documentation and Moore Threads inventory regression coverage

This is the linear-history equivalent of #41 because repository policy disables merge and squash commits and GitHub cannot rebase a PR that contains merge commits. Its Git tree is exactly identical to the fully tested integration head:

aecc9d7a6bb03135501d8388b80ad66670c25a9f

Validation

• 34 passed on A100
• Python compile checks, shell syntax checks, and git diff --check pass
• CUDA 13/A100: full memory, STREAM, L1, and L2 suite passed; final-tree STREAM passed
• MetaX C550, Iluvatar TG-V200, Hygon Z100L, Moore Threads S5000, and Cambricon MLU590: build plus short STREAM passed
• Ascend 910B4: build and short STREAM completed successfully; the development container returned its known exit code 137 after printing completion
• Moore Threads inventory confirmed on hardware as 8 x MTT S5000

[baominghelly]

Superseded by #44. The replacement PR uses the repository-compliant hardware/platform-integration branch and includes the requested design corrections: no tracked implementation plan, hardware platform constants moved to the hardware module, and Dispatcher restored to the single cudaUnified framework.