Critical: Add tests to autograd, nn_primitives and nn

[mratsim]

Autograd, nn_primitives and nn are in master. To bring them to the same standard as the core tensor library, they need tests:

• unit tests for individual pieces (like derivative of linear: Weight * input + bias)

  • Derivatives and cost functions in particular are critical.

• Full pipeline test. (learning XOR, or small dataset like Iris / dogs vs cats). Note: loading those datasets should not depend on modules that depends on Arraymancer.

  • Tests should catch convergence regression
  • A performance and memory benchmark would be nice to have to profile the library regularly. Continuous integration would be terrific but a file in "benchmarks" folder with a history of commit + CPU/GPU + compilation flags would be a great start.

• End-to-end integration tests with IO libraries like arraymancer-vision (and later csv loading, etc).

  • Make sure we don't silently break Arraymancer for them
  • Make sure they don't break assumptions we rely on, like color channel being in CHW order.