Add `alpha` argument to `MixUpSoftLabelDataset`

[Takuya Akiba]

In the original paper of mixup (https://arxiv.org/pdf/1710.09412.pdf), λ (mix_ratio in the code) is sampled from Beta(α, α), where α is a hyperparameter. On the other hand, in ChainerCV's MixUpSoftLabelDataset, mix_ratio is sampled from uniform distribution (which is the same as α=1 case). It might be more useful if we could specify the value of α.

chainercv/chainercv/datasets/mixup_soft_label_dataset.py

Line 67 in 6ce2d47

 mix_ratio = np.random.random()