Support fp16 data type for more operations and fix the minor errors.  (Don't use this! It's an incomplete version)

Support fp16 data type for more operations and fix the minor errors.  (Don't use this! It's an incomplete version)

Use tensor as the bottom implementation and support plain scalar.
Note: DO NOT support the backward for ReduceSumAll function.

1. Move the New/Initial/Del Tensor functions from XTensor.* to XElement.*
2. Move IsSameShaped functions from XTensor.* to core/shape/IsSameShaped.*
3. Move CheckData functions from XTensor.* to core/utilities/CheckData.*

2. Add MemManager class to manage the memory;
3. Use the new version XList;
4. Add a new interface to create Tensor;
5. Apply MemManager to FNNLM.