Refactor code CUDA and Native backend to match #autumnai/collenchyma/62 that
provides enchanced memory management and syncronization. Since memory
management is now automatic, `*_plain` variants of functions are removed.

BREAKING CHANGE: *_plain versions of API functions are removed, arguments of
their counterpart functions may have changed in mutablity.

REFERENCE: autumnai/collenchyma#37, autumnai/collenchyma#62

Refactor tests to be generic on backend and element type, use generic
functions to fill input vectors and check outputs. Use macros to instantiate
concrete tests for Native/Cuda and f32/f64. Tests are now easier to read
and 5-10x shorter.

Add randomly generated test vectors for relu, sigmoid, tanh, softmax and
log softmax. LRN, pooling and convolution use old test vectors. Convolution
test fails the same way it failed before restructure.

Convert and use macros to make benches definitions more compact. Implement
benches for Cuda, though they take ages to complete.

Since benches are moved in main source tree, feature flag "unstable" is used
to conditionally compile them.

BREAKING CHANGE: use cargo flag "unstable" to compile benches.