[Breaking] Accept multi-dim meta info.

[trivialfis]

This is the last part of the original PR for extending meta info to multi-dimension arrays. This PR changes base_margin into a 3-dim array, with one of them being reserved for multi-target classification. Also, a breaking change is made for binary serialization due to extra dimension along with a fix for saving the feature weights. Lastly, it unifies the prediction initialization between CPU and GPU. After this PR, the meta info setter in Python will be based on array interface.

Original description

• Extend the array interface class to multi-dim.
• Define tensor storage class as an extension to host device vector.
• Move some calculation of tensor view to compile time for better checks.
• Use them for base margin.
• Use array interface for Python meta info.
• Unify InitPredOutput function from both predictors.

Motivation

These are basic infrastructures for multi-target model training. The type erased array interface class is used to consume user inputs and typed tensor/tensor_view is used for internal algorithms. For the typed tensor, it makes sense to consider existing implementations instead of rolling out our own. Here are a few things that we need to consider and I haven't found a candidate yet. Firstly the tensor needs to be trivially passed into CUDA kernel. Secondly, we need to have control over its memory allocation for both host and device. Lastly, it should work with any type of data instead of numeric-only since we have custom gradient and might use it to manage the grid of trees (model slicing).

The implementation here satisfies all the above items and features efficient indexing and slicing, along with compile-time dimension safety checks.

To-dos:

• Unify device side meta info initialization.
• More tests.
• JVM packages native code.
• Require latest __cuda_array_interface__.
• Update binary format.
• Python test for f-contiguous input.

[trivialfis]

Now let's see how many compilers are failing. ;-)

[trivialfis]

Hmm??

C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.1\include\thrust/detail/cpp11_required.h(23): fatal error C1189: #error: C++11 is required for this Thrust feature; please upgrade your compiler or pass the appropriate -std=c++XX flag to it. [C:\Users\Administrator\workspace\xgboost-win64_PR-7405\build\src\objxgboost.vcxproj]

[trivialfis]

Note to myself:

• Some tests are failing because the getters of DMatrix don't understand ndarray.
• Some tests are failing because they pass incorrect shapes of meta info, like np.empty(0, 0) for feature weights.
• Some code was written assuming cuda is the only place array interface is available (jvm gpu).

[trivialfis]

Will start extracting components for reviews.