Race the eleven rows that were excluded, by making them do the job

[godofecht]

What this closes

Eleven of the 203 exported estimators were not ranked. Six of those have no
scikit-learn counterpart and never will. The other nine were excluded because
of what our own implementation did, and that is what this fixes.

bucket	before	after
raced	172	197
simplified stand-ins	4	0
doing part of the job	5	0
no counterpart	6	6

The four that were stand-ins

Each said so in its own comments, so the registry timed it and showed the
number without ranking it, and the page carried a paragraph explaining why
three of the four would otherwise have been its widest wins. They were wide
because they were not doing the work.

row	was	is	Flow	scikit-learn
spectral_coclustering	row sums bucketed between min and max	Dhillon: normalized adjacency, singular vectors after the trivial one, one k-means over rows and columns together	0.04 ms	4.73
spectral_biclustering	row and column means bucketed the same way	Kluger: Sinkhorn-Knopp, vectors ranked by how well a one dimensional k-means fits them, rows and columns clustered separately	0.15	30.60
elliptic_envelope	random subsets scored by the product of the covariance diagonal	FastMCD concentration steps, determinant from a Cholesky factor, full Mahalanobis distance in predict	0.80	18.36
nca	a distance that squared each product before summing, and a gradient missing its expectation term	the actual objective and its gradient	251.9	1167.1

Those four ratios are from the gate run on this branch, which reported 192 of
192 with the simplified bucket empty.

The five that did part of the job

A second entry point beside each one does the whole thing.

• voting_classifier_full_fit, voting_regressor_full_fit: fit the ensemble
  from the data. Three trees on both sides, because a vote of one is not what
  either library is being asked for.
• stacking_classifier_full_fit: cross-validates the base estimators to build
  out-of-fold meta-features, fits the meta-learner, refits the base estimators.
• self_training_classifier_full_fit: refits its base estimator every round.
• incremental_pca_fit: walks the design in batches, sized as scikit-learn
  sizes its own.

row	Flow	scikit-learn
incremental_pca	0.010 ms	1.01
voting_classifier_full	0.053	2.23
self_training_classifier_full	0.266	19.83
stacking_classifier_full	0.529	16.07
voting_regressor_full	1.184	3.74

The older functions stay, each with a reason naming the full fit that replaces
it for racing purposes.

A bug racing them found

IncrementalPCA did not work. Its partial fit accumulated the same per-feature
quantity for every component, so every component came out as the same vector,
and the step it called a power iteration multiplied a component by itself
rather than by a covariance. It keeps a running scatter matrix now, exact
across batches once the shift between two means is accounted for, and takes
its components from the leading eigenvectors of it. On the fixture the leading
component carries a spread of 640 against the second's 4, where the two were
previously identical to the digit.

Nothing caught this before because the estimator was never raced and no test
asserted on its output.

Tests

Three new files, all checking behaviour:

• both biclusterings have to recover a planted block structure, with the row
  block and the column block it belongs with agreeing;
• every planted outlier has to be found with no false alarms;
• the learned projection has to separate the classes better than the identity;
• the ensembles and the stack have to classify or regress their fixture, self
  training has to label the unlabelled rows and get them right, and the
  batched decomposition has to see every row and rank its components.

Each was verified as a real gate by inverting an assertion and confirming a
non-zero exit.