feat: add Cambricon hardware benchmarks - #49

Summary

• add the native Cambricon memory, STREAM, NRAM, and L2 cache benchmark suite
• integrate Cambricon platform aliases, runtime detection, build selection, and result parsing into the existing hardware adapter
• add focused adapter tests and update hardware benchmark documentation
• correct the per-core NRAM workload and publish it as hardware.nram_bandwidth
• use distinct pinned host buffers for bidirectional transfers and size all buffers to the largest 1 GiB sweep case
• configure CMake to compile and link with a single cncc invocation
• centralize NRAM alignment and chunk calculation across all seven kernels
• consolidate the four STREAM warmup, notifier, and statistics paths without changing output
• rebase onto the latest master while preserving the existing Ascend and InfiniOps integrations

Testing

• tested commit: 33f7c424a20e915e73ef1d4369705865ba1a0a10
• rebased onto master: 11269f2c31b191cc35c67ade0abd0a4798e386a6
• complete local test suite: 78 passed
• Black 23.9.1: passed
• Python compileall: passed
• git diff --check: passed
• clean CMake configure and verbose build in accelerator-dev/cambricon:latest: passed
• canonical build.sh native build: passed
• MLU590-M9C physical device 0 on its local NUMA node 0: full adapter path passed
• benchmark command: mlu_perf_suite --all --device 0 --iterations 2 --array-size 1048576
• validated all 10 expected metrics:
  • hardware.mem_sweep_h2d
  • hardware.mem_sweep_d2h
  • hardware.mem_sweep_d2d
  • hardware.mem_sweep_bidirectional
  • hardware.stream_copy
  • hardware.stream_scale
  • hardware.stream_add
  • hardware.stream_triad
  • hardware.nram_bandwidth
  • hardware.gpu_cache_l2
• 1024 MiB H2D/D2H/D2D/bidirectional: 27.57 / 27.57 / 348.68 / 36.04 GB/s
• NRAM result: 15402.2 GB/s over 41.6 ms at 1.9 TFLOPS
• normalized 175-line benchmark stdout signature matches the previously validated head be7f00fb
• native log SHA256: B4E5CF9E398E49970FDC893864D4B2BC076883DFE0868B490EA2A496071412CB

Host-device bandwidth was measured with the process pinned to the accelerator-local NUMA node to avoid remote-NUMA host-memory placement. The stdout comparison normalizes only dynamic numeric values and horizontal padding; headings, line structure, table labels, metric order, fields, and units are unchanged.