Add per-observable error summary statistics to Benchmark

[simsaidan]

Pointwise and norm-based errors (L∞, L2) are useful, but benchmarks also need simple summary statistics that are easy to read in logs, README examples, and future CSV exports — especially when comparing many observables at once.