Add Power-aware Benchmarking Support

[jiannanWang]

This PR adds power-aware benchmarking for GPU kernels in BackendBench.
It introduces:

• PowerManager: Collects GPU power, temperature, and frequency data during benchmarks.
• do_bench_power: Benchmarks kernel execution while measuring energy consumption.
• PerformanceTestResult: Now records total energy used per test.
• Power plots: Generates CSV and visualizations for power, temperature, and frequency.
• Dependencies: Adds nvidia-ml-py and matplotlib for power monitoring and plotting.
  This enables precise measurement of energy usage for GPU kernels, supporting energy-efficient optimization.

Example (check the "total_energy" entries):

  {
    "op_name": "add.Scalar",
    "args": "((T([100, 1], i64), 1,), {})",
    "speedup": 1.0037302273210138,
    "total_energy": 0.0011545763355,
    "benchmark_time_ms": 0.006022225360552894,
    "reference_time_ms": 0.006044689630126131,
    "error_msg": "",
    "successfully_ran": true,
    "test_type": "performance"
  },
  {
    "op_name": "add.Tensor",
    "args": "((T([128100, 1536], f16), T([128100, 1536], f16),), {})",
    "speedup": 0.999044778094025,
    "total_energy": 0.13822618472830014,
    "benchmark_time_ms": 0.5277284108675443,
    "reference_time_ms": 0.5272243131290782,
    "error_msg": "",
    "successfully_ran": true,
    "test_type": "performance"
  },
  {
    "op_name": "add.Tensor",
    "args": "((T([256, 1024, 1024], f16), T([256, 1024, 1024], f16),), {})",
    "speedup": 1.0008341558163996,
    "total_energy": 0.1871731542646,
    "benchmark_time_ms": 0.7144741316636404,
    "reference_time_ms": 0.7150701144162346,
    "error_msg": "",
    "successfully_ran": true,
    "test_type": "performance"
  },
  {
    "op_name": "add_.Tensor",
    "args": "((T([128, 512, 28, 28], f16), T([128, 512, 28, 28], f16),), {})",
    "speedup": 1.0001593311388208,
    "total_energy": 0.03350459256860003,
    "benchmark_time_ms": 0.1444975220835301,
    "reference_time_ms": 0.14452054503828043,
    "error_msg": "",
    "successfully_ran": true,
    "test_type": "performance"
  },

[msaroufim]

So the way I'd go about this before adding a utility is to take a couple of pytorch operators on H100 and B200, plot their runtimes in a loop and track the temperature and power draw respectively and see what we can learn

[jiannanWang]

So the way I'd go about this before adding a utility is to take a couple of pytorch operators on H100 and B200, plot their runtimes in a loop and track the temperature and power draw respectively and see what we can learn

Sounds great! I'll do the experiment.