self-document the benchmark's setup

[stas00]

This PR extends the report to include a header that includes important details about the benchmark so that the results are self-documenting/reproducible.

It also adds an optional --notes cli arg if the experimenter needs to add additional notes to be included in the report.

Now it should be much easier to reproduce the report.

Here is an example:

$ python mm_flops.py -m 1 -n 2048 -k 2048 --output_file=mm_m_range.txt --notes="numa_balancing=0"

Benchmark started on 2024-02-17 02:20:50

** Command line:
/env/lib/conda/py39-pt23/bin/python mm_flops.py -m 1 -n 2048 -k 2048 --output_file=mm_m_range.txt --notes=numa_balancing=0

** Platform:
Linux example 5.15.0-1004- #4-Ubuntu SMP Fri Jan 26 21:18:52 UTC 2024 x86_64 x86_64

** Critical component versions:
torch=2.3.0.dev20240216+cu121, cuda=12.1, nccl=(2, 19, 3)

** Additional notes:
numa_balancing=0

--------------------------------------------------------------------------------


Elapsed time for 1x2048x2048: 0.000
Throughput (in TFLOP/s) for 1x2048x2048: 1.502
--------------------------------------------------------------------------------

You can later add other component versions that you think are important

[stas00]

hmm, probably should try to detect nvidia-smi or rocm-smi and dump those as well, right? as it would also be useful to have in the log file to know exactly what the file was generated on, including the driver versions.

anything else?