fix bug at mha, MaskGenerator; improve ckpt_utils.py

[flymin]

1. fix bug in mha when mask=None.
2. fix bug in MaskGenerator. If the video is short, any type of mask has chances to mask all.
3. use logger in load_checkpoint.

[flymin]

Regarding 1, another solution is to change

q = self.q_linear(x).view(1, -1, self.num_heads, self.head_dim)

to

q = self.q_linear(x).view(B, -1, self.num_heads, self.head_dim)

if mask is None. Current implementation on the main branch assumes mask is never to be None.