Fix bugs in Scholar.Linear.IsotonicRegression

[RicardoSantos-99]

• A decreasing fit can come back non-monotonic. contiguous_isotonic_regression/4
  writes each pooled value back over the block it covers, but that pass stops at
  max_size, which is only where the valid range ends when fitting increasing.
  Fitting decreasing reverses the input, so the placeholder slots left by tied x
  values move to the front and the range ends at y_size instead. Part of the
  last block keeps its unpooled value:
  fit([0, 0, 1, 1, 2], [4, 4, 5, 5, 3], increasing: false) returned thresholds
  [4.0, 4.5, 3.0], which rises where the caller asked for a decreasing fit.
• A zero weight poisons the whole fit. A sample with weight zero still opens
  or joins a group, so a group made only of those divides by zero. The NaN then
  spreads through the pooling: sample_weights: [1, 0, 1, 1, 1] returned
  thresholds of -infinity and predicted NaN everywhere. scikit-learn masks
  non-positive weights out before fitting.
• out_of_bounds was never read. It appears once in the whole module, in the
  schema. predict/2 always clipped, so :nan, the documented default, behaved
  exactly like :clip and the default behaviour did not exist.
• preprocess/1 raises on models with fewer than three thresholds. The
  duplicate trim slices y[1..-2//1], which needs three. Constant targets or a
  two point fit produce two, and the range 1..0 raises.
• cutoff_index goes stale. The trim drops thresholds but the index keeps
  pointing into the untrimmed tensors, so running preprocess/1 on its own
  output slices past the end and raises.
• fit/3 rejects the column shape its own check allows. check_input_shape/1
  documents {n, 1} as valid and predict/2 accepts it, but fit/3
  destructures the shape as {n_samples} and raises a MatchError first.
• increasing: :auto follows a least squares slope. One outlier flips it.
  scikit-learn uses the sign of Spearman's rho, which reads ranks and does not
  move with the size of an outlier. On x = 1..10 with y = 1..9 then -100,
  the slope is -5 and picks decreasing while rho stays at 0.45 and picks
  increasing.

[josevalim]

💚 💙 💜 💛 ❤️