Fix fast ABOD counting each training point as its own neighbor

[raashish1601]

All Submissions Basics:

• Have you followed the guidelines in our Contributing document?
• Have you checked to ensure there aren't other open Pull Requests for the same update/change?
• Have you checked all Issues to tie the PR to a specific one? (No existing issue; ABOD weighting schemes / not matching the definition? #522 is about the ABOF weighting, which this PR does not touch.)

All Submissions Cores:

• Have you added an explanation of what your changes do and why you'd like us to include them?
• Have you written new tests for your core changes, as applicable?
• Have you successfully ran tests with your changes locally?
• Does your submission pass tests, including CircleCI, Travis CI, and AppVeyor? (CI will run on this PR.)
• Does your submission have appropriate code coverage? The cutoff threshold is 95% by Coversall.

In fast ABOD, _fit_fast finds the neighbors of the training points with self.neigh_.kneighbors(self.X_train_, ...). Querying with the training data returns each point as its own nearest neighbor, and _calculate_wocs then skips every pair that contains the point itself, so a training score is computed from only n_neighbors - 1 real neighbors. decision_function uses n_neighbors neighbors for new points, so decision_scores_ (and threshold_ / labels_, which come from them) are on a different basis than the test scores. As a consequence, fast ABOD with n_neighbors = n_train - 1 also does not match the default (all points) method.

The fix calls kneighbors(n_neighbors=..., return_distance=False) without X, which in scikit-learn excludes each training point from its own neighbors (the same approach KNN uses). Only fast ABOD training scores change; decision_function and the default method are unchanged.

Tests in pyod/test/test_abod.py:

• test_fast_mode_train_scores_use_k_other_points: each training score equals -_calculate_wocs over its n_neighbors nearest other training points.
• test_fast_mode_with_all_neighbors_matches_default: with n_neighbors = n - 1, fast and default give the same training scores.

Both fail on the current code and pass with the change. _numpy_abod_scores, the reference added by #766 for the dtype tests, queried the training points with the training data too (it followed the current convention), so it now also leaves each training point out of its own neighbors. pyod/test/test_abod.py passes locally (58 tests).