feat(fitting): support continuous [0,1] responses in IRT fitting methods (fixes #7)

[sparkyluvscode]

Summary

IRT fitting methods (MLE, EM, JML) crashed on continuous [0, 1] response data because the default bernoulli_nll loss routes through torch.distributions.Bernoulli, whose domain validation rejects non-{0, 1} values. Users with partial-credit or soft-label scoring data had no option besides BetaRasch, which assumes a specific distributional shape that rarely fits uniform-ish [0, 1] data.

Adds cross_entropy_nll to _losses.py, which computes the same formula as bernoulli_nll via F.binary_cross_entropy without domain validation. All three fitting methods now auto-detect continuous responses and select the appropriate loss when loss_fn is not explicitly provided. Binary {0, 1} data continues to use bernoulli_nll with no behavior change.

Fixes #7

Tests

• TestCrossEntropyNLL: accepts continuous targets, matches bernoulli_nll on binary, confirms bernoulli_nll rejects continuous, gradient flows.
• TestMLEFitContinuous: Rasch fits continuous data via MLE (Adam + LBFGS).
• TestEMFitContinuous: Rasch fits continuous data via EM.
• TestJMLFitContinuous: Rasch fits continuous data via JML with loss decrease.

All 56 existing fitting tests unaffected; 294 total tests pass.