Promote dev to main for 0.6.3

[IvanChernyshov]

This PR promotes dev to main for the 0.6.3 release and rolls the released version forward from 0.6.1 to 0.6.3.

It includes the cumulative 0.6.2 and 0.6.3 changes.

Highlights since 0.6.1

Public power-fit problem/result API

• Adds pyvoro2.powerfit.build_power_fit_problem(...) with stable PowerFitProblem, PowerFitBounds, and PowerFitPredictions objects for external evaluation and solver experiments.
• Adds pyvoro2.powerfit.build_power_fit_result(...) so externally computed candidate weights can be repackaged as standard PowerWeightFitResult objects.
• Adds PowerFitObjectiveBreakdown and PowerWeightFitResult.objective_breakdown for explicit hard-bound and per-penalty summaries.
• Makes exposed problem-definition arrays read-only to align with frozen dataclass semantics.
• Makes PowerWeightFitResult.status open-ended and pair it with status_detail so externally packaged results can preserve solver-specific termination details.

Diagnostics and reporting

• Adds PowerWeightFitResult.edge_diagnostics with edge-space arrays and summaries, including alpha, beta, z_obs, z_fit, algebraic residuals, and weighted/unweighted algebraic inconsistency metrics.
• Extends fit reports and per-constraint record exporters to serialize the new algebraic edge diagnostics alongside existing measurement-space predictions and residuals.

Solver and numerical improvements

• Updates solver='admm' to split on the predicted measurement variable y = beta + alpha (w_i - w_j) instead of raw weight differences.
• Warm-starts ADMM from the quadratic analytic fit when that warm start is identifiable.
• Fixes the native Huber/ADMM path so representative sparse-outlier cases no longer stall or blow up.

Internal consistency and structure

• Centralizes prediction, algebraic diagnostics, hard-bound conversion, gauge canonicalization, and objective evaluation in pyvoro2.powerfit.problem.
• Moves shared public power-fit dataclasses to pyvoro2.powerfit.types.
• Moves pure public weight/radius conversions to pyvoro2.powerfit.transforms.
• Routes the native solver, active-set layer, and reporting layer through the shared problem/evaluation machinery to eliminate internal formula drift.

Regression coverage

• Adds a medium-size robust-fit regression test for the native Huber/ADMM path.
• Adds an advanced problem-API regression suite covering export, read-only arrays, result repackaging, and status-detail reporting.