feature(wrh): add adaptive batch size for transition

[ruiheng123]

1. Each env has different steps per episode. We improved the sampling batch size distribution according to the collected steps in the buffer. Env with more collected steps can obtain larger batch size for training.
2. Meanwhile, we use softmax with temperature T param to keep the difference between different batch sizes from being too large.