Implemented NSIM similarity measurement.

[zond]

• Implemented NSIM (DOI:10.1016/j.specom.2011.09.004).
• Replaced the geometric norm calculation across time and frequency axes with '1 - NSIM' when computing Zimtohrli distance.
• Removed the redundant cpp/zimt/README.md.
• Refactored the dataset management tools to enable optimization without saving the intermediary results.
• Implemented simple simulated annealing optimization for some relevant Zimtohrli parameters.
• Redid the MOS mapping so that Zimtohrli MOS scores are still on the same scale as before the NSIM change.

Before:

Score type	MSE	Min score	Max score	Mean score
ViSQOL	0.09	0.52	0.80	0.62
PESQ	0.10	0.48	0.84	0.61
Zimtohrli	0.10	0.56	0.74	0.59
CDPAM	0.12	0.44	0.73	0.57
PARLAQ	0.15	0.47	0.78	0.54
PEAQB	0.18	0.28	0.85	0.51
DPAM	0.28	0.19	0.69	0.42
WARP-Q	0.28	0.07	0.78	0.45
GVPMOS	0.32	0.01	0.78	0.40

After:

Score type	MSE	Min score	Max score	Mean score
ViSQOL	0.09	0.52	0.80	0.62
Zimtohrli	0.09	0.59	0.72	0.60
PESQ	0.10	0.48	0.84	0.61
CDPAM	0.12	0.44	0.73	0.57
PARLAQ	0.15	0.47	0.78	0.54
PEAQB	0.18	0.28	0.85	0.51
DPAM	0.28	0.19	0.69	0.42
WARP-Q	0.28	0.07	0.78	0.45
GVPMOS	0.32	0.01	0.78	0.40

See https://github.com/google/zimtohrli/pull/64/files?short_path=4640799#diff-46407991f9df1906835213b977238d969f72c49974a7dda6a9b1276daf199221 for the detailed performance changes.