ENH: Add justify functionality

I used this functionality quite often and I think would fit in numpy. It's to shift values over an invalid value to a certain direction over certain axis.

Note: this code and functionality was provided by Divakar on SO

Reproducing code example:

a = np.array([[1, 0, 2, 0],
              [3, 0, 4, 0],
              [5, 0, 6, 0],
              [6, 7, 0, 8]])

def justify(a, invalid_val=0, axis=1, side='left'):    
    """
    Justifies a 2D array

    Parameters
    ----------
    A : ndarray
        Input array to be justified
    axis : int
        Axis along which justification is to be made
    side : str
        Direction of justification. It could be 'left', 'right', 'up', 'down'
        It should be 'left' or 'right' for axis=1 and 'up' or 'down' for axis=0.

    """

    if invalid_val is np.nan:
        mask = ~np.isnan(a)
    else:
        mask = a!=invalid_val
    justified_mask = np.sort(mask,axis=axis)
    if (side=='up') | (side=='left'):
        justified_mask = np.flip(justified_mask,axis=axis)
    out = np.full(a.shape, invalid_val) 
    if axis==1:
        out[justified_mask] = a[mask]
    else:
        out.T[justified_mask.T] = a.T[mask.T]
    return out

Output

In [473]: a # input array
Out[473]: 
array([[1, 0, 2, 0],
       [3, 0, 4, 0],
       [5, 0, 6, 0],
       [6, 7, 0, 8]])

In [474]: justify(a, axis=0, side='up')
Out[474]: 
array([[1, 7, 2, 8],
       [3, 0, 4, 0],
       [5, 0, 6, 0],
       [6, 0, 0, 0]])

In [475]: justify(a, axis=0, side='down')
Out[475]: 
array([[1, 0, 0, 0],
       [3, 0, 2, 0],
       [5, 0, 4, 0],
       [6, 7, 6, 8]])

In [476]: justify(a, axis=1, side='left')
Out[476]: 
array([[1, 2, 0, 0],
       [3, 4, 0, 0],
       [5, 6, 0, 0],
       [6, 7, 8, 0]])

In [477]: justify(a, axis=1, side='right')
Out[477]: 
array([[0, 0, 1, 2],
       [0, 0, 3, 4],
       [0, 0, 5, 6],
       [0, 6, 7, 8]])

Start by reviewing the provided Python justify implementation and run the reproducing examples against NumPy arrays. Compare the four axis and direction cases with the shown outputs, then determine the appropriate NumPy API and edge-case behavior. Done means the functionality is integrated with documented behavior and tests cover the demonstrated cases.