test: validate FMM numerical correctness

[project-relay]

Adds an end-to-end numerical validation harness before changing the solver implementation.

What changed:

• adds an analytic M2L dipole-orientation check
• compares FMM forces against an independent O(N^2) softened-gravity reference
• covers deterministic uniform and asymmetric two-cluster distributions
• reports RMS and maximum relative force error
• runs the validation through CTest/CI with deliberately generous 10% RMS smoke-test gates

Why:
The current benchmark establishes runtime scaling but does not establish approximation accuracy. This isolates correctness evidence first so any implementation fixes can be measured rather than inferred from the demo or performance graph.

@project-relay