Multiple changes:

1. use unique_ptr instead of shared_ptr, and fix a bug in destructor of TrtEngineOp where it did't reset the shared_ptr but a copy of it
2. fix the include order
3. shorten the reference to tensorflow::tensorrt::xxx
4. remove some code that sets something which will be overwritten later
5. fix format, including: function signature, variable names, const reference, etc
6. remove some deadcode
7. add a lot of comments and TODOs
8. in TrtEngineOp, replace the map of allocators with a single unique_ptr
9. in TrtEngineOp, remove parameter ignore_dim_change from GetEngine(), since it always uses member fixed_input_size_