Add conditioning diagnostics for multimode fits

[maiconburn]

What this adds

A small module ringdown.conditioning with two utilities meant to be run before or alongside a fit:

• mode_gaps(modes, chi): pairwise distances between the mode frequencies of a fit model in the complex plane (dimensionless, or in 1/s when a mass is given).
• conditioning_report(modes, chi): for each mode pair, the gap and the error-amplification factors for three different estimation tasks, sorted by gap, with the smallest-gap pair flagged.

The amplification exponents encode how hard each task gets as two modes approach each other:

task	scaling
amplitudes, frequencies fixed by (M, chi)	1 / gap
frequencies themselves	1 / gap^2
amplitudes with free frequencies	1 / gap^3

Example numbers

GW150914-like model (220 + 221), chi = 0.69:
  gap = 0.165   1/gap = 6.1    1/gap^2 = 37    1/gap^3 = 224

(2,2,5) + (2,2,6) near their avoided crossing, chi = 0.897:
  gap = 0.0667  1/gap = 15     1/gap^2 = 225   1/gap^3 = 3374

Same pair away from the crossing, chi = 0.70:
  gap = 0.178   1/gap = 5.6    1/gap^2 = 32    1/gap^3 = 179

The first line is one way to see, quantitatively, why overtone amplitudes with free frequencies are so fragile even for a loud event, and the second why nobody separates n = 5 from n = 6 near the crossing.

Where the exponents come from

The 1/gap scaling for fixed-frequency amplitudes was measured on SXS waveforms across the (2,2,5) and (2,2,6) avoided crossing, and the full task hierarchy (1, 2, 3) was measured in a controlled synthetic study; both live at https://github.com/maiconburn/recoverability-criticality (DOI 10.5281/zenodo.22156019). The 1/gap^2 frequency case matches the Fisher forecast scaling of arXiv:2605.16199, and the 1/gap^3 case matches the classical Prony-type super-resolution scaling. The exponents are stated up to mode-dependent prefactors, so the factors are most meaningful when compared between pairs or across spins.

Tests

Four tests included, anchored on the known (2,2,5)-(2,2,6) crossing (gap 0.0667 at chi about 0.897). On my machine tests/test_conditioning.py and tests/test_qnms.py pass together (21 tests); the data-dependent test files need the download scripts and were not touched.

Happy to adjust naming, API, or placement, or to wire the report into Fit if that would be useful.