*fix bug: segmentation fault at test case

[joddiy]

*fix bug:
- allow reshaping on an empty tensor then give it data will invoke segmentation fault.
*how to fix:
- when reshape, add a check requiring the original shape's volume must be same with new one.
- fix corresponding test case(test_mse, test_cross_entropy, test_tensor_math)

[nudles]

The tranpose operation should return a new tensor that shares the data with this tensor if possible; otherwise return a new tensor that copies the data.

[joddiy]

The transpose is another problem, I think yisen will submit a new PR(otherwise I'll simbit a new PR to copy that).

[nudles]

maybe the current transpose implementation is ok. T() and Tensor::Transpose() change itself. The global transpose return a view if possible. No need to change now.

[joddiy]

But the current transpose implementation cannot pass the test case. The test case
Tensor o = t.T(); EXPECT_EQ(t.shape()[0], o.shape()[1]); EXPECT_EQ(t.shape()[1], o.shape()[0]);
seems the transpose func needs to return a new transposed tensor, rather than change itself.