Add top-label histogram binning as a post-hoc confidence calibrator

[Yasas-Sri]

Closes #8.

Adds anyjev/calibrate/binning.py: top-label histogram binning (Zadrozny and Elkan, 2001) as a standalone post-hoc confidence calibrator, with fit / apply / to_dict / from_dict. It works on any [N, K] probabilities AnyJev returns, fit on human labels or the model's own answers after reasoning.

• Equal-mass bins over top-1 confidence, each mapped to its measured top-1 accuracy. Bins under min_count are merged into a neighbour.
• The argmax is kept. The other options are scaled by one common factor, and the new top-1 is floored just above the rescaled runner-up, so a low bin value can't let the runner-up overtake it. Accuracy is exactly the input's.
• No defaults on the fields: no bin table leaves probabilities unchanged, so HistogramBinning() without fitting raises.
• Outputs are clipped to at least 1e-12 and renormalised, so a bin that was always right gives no exact 0 or 1, and a later log stays finite.

Tests (tests/test_binning.py): exact per-bin values on constructed sets, argmax kept and rows summing to 1 on random data (K = 2, 4, 20), the runner-up case, confidence 1.0, no exact 0/1 outputs, to_dict / from_dict round trip, bad input, and a held-out ECE check on an overconfident model (with a local ECE helper)

ruff check anyjev scripts space tests and pytest -q pass locally.