Add sweep_tunable_mask_size/recommend_target_impact_share

[edkerk]

select_tunable_mask's target_impact_share picks a comparable share
of leverage on any model, but the same share does not sit at a
comparable point on the fit curve -- two models with similarly shaped
leverage curves can have fit curves that diverge sharply. The only way
to tell is to tune at a few candidate sizes and look at where the
improvement stops.

This adds two functions to geckopy.kcat_tuning.evotune:

• sweep_tunable_mask_size tunes at each of a list of
  target_impact_shares (and seeds) and reports the resulting fit per
  point, reusing one screen across all of them.
• recommend_target_impact_share picks the smallest share past which
  further tuning stops paying for itself, using two independent
  checks against a later point's best result: a relative-improvement
  tolerance, and, where at least one of the two points compared has
  more than one seed, a check against their own local seed spread
  (not a sweep-wide pooled one, since how noisy a comparison is can
  differ across the sweep).

Along the way: MaskSizePoint.rmse takes min(rmse_trace) per seed
rather than the final entry, since rmse_trace only advances at a
step that improved the (possibly prior-penalised) objective, per
cmaes_kcat_tuning's own docstring -- with the default nonzero
prior_penalty_weight the last entry is not guaranteed to be the best
one recorded.

sweep_tunable_mask_size's default target_impact_shares now starts
at 0.3 (not 0.5) and its default seeds is (0, 1) (not a single
seed), so a default call's noise-floor check has something to compare
against instead of silently degrading to the relative check alone.

Built and validated against real sweeps on both the E. coli and yeast
models worked with in this branch's originating session; matches
manual findings on both.

New tests in tests/test_mask_size.py; full suite passes
(1448 passed, 1 deselected, 1 xfailed).