C++ API and architecture refactor

[kparichay]

This issue is to track the re-factor of the exposed headers and the C++ API. This will update the NNTrainer architecture to the class diagram finalized with @jijoongmoon.

• [x] Update internal implementation to use only the methods exposed by those headers and not an internal implementation
  • [x] Remove all friends of layers
  • [x] Manage input/output properties better for layer - track at #1454
  • [x] add unittests for ensureName related to #1009
  • [x] refactor graph to move graph node related members from layers
  • [x] #1029
  • [x] consider removing adj lists once compile is done to save some space. - #1049
  • [x] move layers connection related properties out to graph - #1142 #1182
  • [x] tensor constructor to support initialization - https://github.com/nnstreamer/nntrainer/pull/1031#discussion_r597551354 - pointed out by @myungjoo - #1452
  • [x] refactor model loader to allow custom properties #1032
  • [x] prepare custom loss #1256
  • [x] many more
• [x] Update headers and their interfaces
  • [x] Make c++ api equivalent to c API - done with #1050
  • [x] C++ API must be superset of https://github.com/nnstreamer/nntrainer/pull/1010#discussion_r594866783 or the nntrainer filter must be updated to accommodate with the API. - #1461
  • [x] Make c api directly use c++ API - #1050
  • [x] divide *_internal to *_devel which will be just an interface and *_impl which will be the default implementation. For example, Layer <- Layer_devel <- Layer_impl. - #1426
• [ ] Update packaging and applications correspondingly #1450

[taos-ci]

:octocat: cibot: Thank you for posting issue #986. The person in charge will reply soon.