feat: add ItemBank for storing and reusing calibrated items

[kholdrex]

Why

Fit results are positional arrays tied to one in-memory model. An app that wants to work with new respondents (for example, a Rails assessment app) has to either recalibrate or invent its own storage for item parameters. This is step 2 of the roadmap and the base for the next step: scoring new respondents (EAP/MAP) against fixed items.

What

IrtRuby::ItemBank is an immutable, id-keyed set of calibrated items. It uses the logistic metric without a scaling constant, the same as R's mirt: P = c + (1 - c) / (1 + exp(-a (theta - b))).

bank = IrtRuby::ItemBank.from_fit(results, item_ids: %w[q1 q2 q3 q4], metadata: { source: "spring pilot" })
File.write("item_bank.json", bank.to_json)
bank = IrtRuby::ItemBank.from_json(File.read("item_bank.json"))
bank.fetch("q2")             # Item: id, a, b, c
bank.probability("q2", 0.5)

• Versioned JSON format. format: "irt_ruby.item_bank", format_version: 1, model, items {id, a, b, c}, optional ability_distribution: {mean, sd}, and free-form metadata. Unknown formats and newer versions are rejected.
• from_fit. Infers the model from the result keys. Records provenance (gem version, scale convention, converged, iterations) and the calibration sample's ability distribution. It never invents values for estimated parameters. Items with nil estimates raise an error unless exclude_unestimated: true is passed, in which case they are listed under metadata["unestimated_items"].
• Per-model validation. Rasch requires a = 1 (stored exactly), c = 0. 2PL requires c = 0. 3PL requires 0 ≤ c < 1. Numbers must be finite and ids must be unique non-empty strings.
• Immutability and equality. The bank, its items and its metadata are deep-frozen and defensively copied. Floats round-trip through JSON exactly. ==, eql? and hash are consistent.
• README. New "Item Banks" section documenting the format, plus an export recipe for R's mirt: it uses coef(..., IRTpars = TRUE, simplify = TRUE) and jsonlite with digits = I(17) for exact doubles, and exports the latent mean and variance.

Verification

• 326 specs pass on Ruby 2.6.10 (json 2.1), 3.3 and 4.0.6. RuboCop 1.92.0 is clean.
• New spec/irt_ruby/item_bank_spec.rb covers building from all three models, JSON round-trip and format, validation, immutability, equality and probabilities. A README example spec is also added.
• The R recipe was checked against the mirt and jsonlite documentation and source, but has not been run in R. A parity check against mirt is the next roadmap step.