Issue 1: Allow scoring function to be a function that does not need to return pvalues

SelectKBest and SelectPercentile both require that the return value be a pair of arrays (scores, pvalues). But this prevents functions that return only one array of scores for each feature, such as

• mutual_info_regression
• mutual_info_classif
• np.var(X, axis=0) (In this case, y is None). lambda X: np.var(X, axis=0) to be precise.

Issue 2: Make a wrapper around functions that score an individual feature

Requested from this comment.

This would be useful for certain scipy correlation functions, such as

• spearmanr
• pearsonr

For example,

def make_feature_scorer(feature_scorer):
    """A wrapper around functions which returns scores for each feature.

    `feature_scorer` should be a function that takes in as input (X, y), where 
    y is allowed to be None.

    Example
    ------------
    >>> from sklearn.feature_selection import make_feature_scorer, SelectKBest
    >>> from scipy.stats import spearmanr
    >>> from sklearn.datasets import make_classification
    >>> 
    >>> X, y = make_classification(random_state=0)
    >>> skb = SelectKBest(make_feature_scorer(spearmanr), k=10)
    >>> skb.fit(X, y)    # Calculates spearmanr for each feature in `X`
    >>> new_X = skb.transform(X)    # Selects the top k=10 features with greatest spearmanr

Note that the make_feature_scorer needs to also handle functions that return a score that might be negative. For example, kendalltau returns values between [-1, 1], where 0 means X has no correlation to the response y.