1. fill_constant support Tensor;
2. fix bug in backward.py: using fill_constant instead of fill_constant_batch_size_like
3. fix bug in ExpandGradOp.
test=release/1.6 (#20521)