Extend nvbench_compare.py with --plot, axis/benchmark filtering, and dark mode

[bernhardmgruber]

Builds on top of #314. Entirely vibe-coded as well.

Example comparing with axis filter:

python python/scripts/nvbench_compare.py -a T{ct}=[F32,F16] -a Elements{io}[pow2]=28 --plot ./pytorch_ublkcpy_B200.json ./pytorch_unroll_2_bif_100.json --dark

With threshold diff and no dark mode:

python python/scripts/nvbench_compare.py -a T{ct}=[F32,F16] -a Elements{io}[pow2]=28 --plot --threshold-diff 0.1 ./pytorch_ublkcpy_B200.json ./pytorch_unroll_2_bif_100.json 

No filters:

python python/scripts/nvbench_compare.py --plot ./pytorch_ublkcpy_B200.json ./pytorch_unroll_2_bif_100.json