Fix BatchNorm Operator error

• Fix batchnorm operator error (8b841e)
• Added updated docs for batchnorm operator (8b841e)

Add BatchNormalization Operator

• Add BatchNormalization Operator. Currently, it supports inference forwarding only.
• In Conv2D operator, use ConvMethod.Img2Col as default calculation method. Since it computes much faster.

Fix CocoaPods installing error

Fix installing via CocoaPods failure error.

Improve Conv2D Peformance

Improvements:

• Improve the Conv2D performance [a95329, 2c113e, c412fa]
• Fix MPS error in matrix multiply operator [57979d]
• Improve Graph API by introducing ForwardGraph [dcfe1f]

Note
This release passed all testing on a local MAC machine. Travis-CI could not pass cause it runs out of time.

Adding guides

This releasee majorly added more guides. Also, modified few APIs involving with slice tensor making.

Changelogs

• Added guides
• Changed APIs in Tensor making slice tensor. [5c7b41f]
• Fixed force downcastign in servera places. [e6f7ac1]
• Added API in SerranoEngine allowing loading user's Metal library. This is for custom operator. [5c28e3a]

Basic APIs and Architecture Ready

This is an alpha development release.

Release note:

Core APIs ready:

• Graph [backward, forward]
• Tensor
• Operator
• Symbol

Operators Ready (Grad compute not ready of some operators):

• Unary operators
• Binary operators
• Reduce operators
• Activation operators
• Matrix multiplication operator
• 2D convolutional operator
• 2D Pooling operators

Next phase goal:

• Graph memory allocation optimization
• Operator grad compute
• Add operators:
  • Depth-wise convolutional operator
  • Batch normalization operator
• Add optimizer:
  • Adam
• Serrano-tool:
  • Convert Keras model