fix: make joint estimation correct and honest about convergence

[kholdrex]

Why

An audit of the estimation code found that results could be wrong or misleading without any signal to the caller:

• 3PL gradients omitted chain-rule factors (8–55% relative error against finite differences), so 3PL optimized the wrong objective. On 1000×20 simulated data, recovered guessing parameters correlated 0.09 with the true values.
• Persons and items without a finite joint MLE (all-correct, all-incorrect, unobserved, perfectly separated or disconnected groups) were returned as ordinary numbers that kept growing with max_iter, or kept their random initial values.
• The latent scale was not identified, and the stopping rule could fire after the learning rate collapsed, far from an optimum. It was still reported as a normal result.
• 1e-15 inside Math.log flattened the objective for badly mispredicted responses. Input and output arrays aliased model state.

What changes

• Estimability: only the largest strongly connected component of the response graph is estimated (correct answer: item → person, incorrect: person → item). Everything else is returned as nil. This guarantees finite estimates for Rasch and for 2PL with bounded discriminations. For 3PL it is necessary but not sufficient, and the README says so.
• 3PL: correct gradients. Excluded all-incorrect respondents keep their limiting log(1 - c) contribution. Guessing-gradient terms are capped so they cannot overflow to Infinity/NaN.
• Scale: results are reported with Rasch abilities centered at 0, and 2PL/3PL abilities standardized to mean 0 / SD 1 with likelihood-preserving item transforms. Internal state is left untouched, so repeated fit calls continue cleanly.
• Convergence: fit returns converged: and iterations:. converged is true only at a stationary point on the reported scale (new gradient_tolerance: option, default 1e-3). The learning rate grows 10% after accepted steps, so it can recover from earlier decays.
• Other: stable log-sigmoid likelihood; defensive copies of input and output arrays.
• Docs: README gains an "Estimation Details" section. The gemspec no longer claims production readiness. The 0.3.0 date in the CHANGELOG is corrected.

Breaking changes (suggest releasing as 0.4.0)

• Result arrays can contain nil for non-estimable persons and items.
• Result hashes have two extra keys, and 2PL/3PL numbers change because of the gradient fix and scale identification.

Verification

• 277 specs pass on Ruby 2.6.10, 3.3 and 4.0.6. RuboCop is clean and the gem builds.
• New spec/irt_ruby/estimation_spec.rb covers:
  • finite-difference gradient checks for every model and missing-data strategy;
  • exclusion of extreme, separated and disconnected groups;
  • scale identification and likelihood preservation;
  • repeated fits and stationarity on the reported scale;
  • 3PL floor terms and overflow;
  • Rasch/2PL parameter recovery.
• Each regression spec fails on the code before its fix.

Known limitations

2PL/3PL joint estimation with plain gradient ascent converges slowly and often reports converged: false on realistic data. That flag is accurate, not a bug. Marginal maximum likelihood is the proper follow-up.