fix: correct off-by-one and tie handling in od p-value computation

[shivamlalakiya]

Fixes #958

What

od/pytorch/base.py:199 and od/sklearn/base.py:148 compute the outlier
p-value as

return (1 + (scores[:, None] < self.val_scores).sum(-1))/len(self.val_scores) \
    if self.threshold_inferred else None

od/pytorch/ensemble.py:114, same author, same subpackage, computes the
same "add one to the count the test score beats" idea but divides by
len(self.val_scores) + 1:

p_vals = (1 + less_than_val_scores.sum(1))/(len(self.val_scores) + 1)

The base.py denominator can push the value above 1. Reproduced through
the public API (0.13.0 wheel, Python 3.11, numpy 1.26.4):

>>> d = GMM(n_components=1, backend='sklearn')
>>> d.fit(X_fit)          # X_fit as in the linked issue
>>> d.infer_threshold(X_cal, fpr=0.5)   # len(val_scores) == 4
>>> d.predict(np.array([[3.5]], dtype=np.float32))['data']['p_value']
array([1.25])

The same line's strict < separately collapses tied scores toward 1/n
instead of 1. Full numbers, and the exact-rational tie-support sweep,
are in #958.

Fix

One line changed per file: < becomes <=, and the denominator becomes
len(self.val_scores) + 1, matching ensemble.py.

return (1 + (scores[:, None] <= self.val_scores).sum(-1))/(len(self.val_scores) + 1) \
    if self.threshold_inferred else None

Verified against the patched files:

>>> s.val_scores = np.full(9, 5.0); s._p_vals(np.array([5.0]))
array([1.])
>>> s.val_scores = np.array([1.0,2.0,3.0,4.0]); s._p_vals(np.array([0.0]))
array([1.])

I searched od/pytorch/tests and od/sklearn/tests for coverage of this
expression and found none, so this is a pure two-line change with nothing
else to update. ensemble.py:114 is untouched; its denominator is already
correct and its strict comparator is a separate, package-wide question
I've left out of this PR (noted in #958).

[CLAassistant]

All committers have signed the CLA.