Minor fix

Use implicit gemm to accelerate conv operator (#15)

Optimize BN backward pass (#14)

Fix no_grad functionality

Add NoGrad class with reference count (#13)

Optimize batch norm by using fused bn

Remove numba requirement

Add sqr and sqrt elementwise ops for cpu backend

Add cpu convolution neural network support (#12)

Fully refactor all ops to decouple with device type (#11)

Fix bug with bias is None

Rename backend ops function names

Decouple tensor with cuda related code

Add backend dispatcher to automatically select backend using device string (#10)

Use np.array2string to represent tensor