[road map] split the core library

[lgarithm]

In a future (not near so far), the core library will be spitted into

a library for basic tensor data type

https://github.com/lgarithm/stdtensor

• should follow the style of std::vector
• should provide the functionality of np.array

a library for computation graph

https://github.com/lgarithm/nn-graph-experimental

• should be generic enough for general dataflow

a library for operators

https://github.com/lgarithm/stdnn-ops

• should contain basic nn operator
• should be easily extendable

The GPU implementation of this library can be just a simple wrapper around cudnn thrust and cutlass.

https://github.com/lgarithm/stdnn-ops-cuda

all of them should provide both C APIs (for embedded system) and C++ APIs.

[lgarithm]

The tensor library should have

• std::tensor (the counter part of std::vector)
• std::tensor_ref (a mutable reference of std::tensor)
• std::tensor_view (an immutable reference of std::tensor, similar to std::string_view)

[lgarithm]

To provide C API, we need to have dynamic types, too. Therefore we gonna to introduce the following 6 categories of data types.

	owner	ref	view
dynamic	tensor_t	tensor_ref_t	tensor_view_t
static	std::tensor<R, r>	std::tensor_ref<R, r>	std::tensor_view<R, r>

where R is the scaler type (float, double), r is the rank of tensor.

[lgarithm]

@luomai compling eager mode in C++ https://github.com/lgarithm/stdnn-ops/blob/dev/examples/example_layer_mlp.cpp