Initial Release: Concurrent Inference Capability

This release provides an inference_engine_t type that encapsulates the state and behavior of a dense neural network with

1. State:
   a. Weights and biases biases gathered into contiguous, multidimensional arrays, 🏋️
   b. Hidden layers with a uniform number of neurons. 🧠
2. Behavior:
   a. A pure infer type-bound procedure that propagates input through the above architecture to produce output. ❄️
   b. An elemental interface for activation functions with one currently implemented: a step function. 🪜
   c. Runtime selection of inference method via the Strategy Pattern:
   • concurrent execution of dot_product intrinsic function invocations or
   • matmul intrinsic function invocations.
     d. Runtime selection of activation functions: currently only a step function is implemented to support the unit tests.
3. Unit tests that
   a. Read and write neural networks to files, 📁
   b. Construct a network that implements an exclusive-or (XOR) gate in the first hidden layer followed by a second layer with weights described by the identity matrix so that the second layer serves a pass-through role. 🚧
4. Examples
   a. Concurrent inference on multiple independent neural networks encapsulated in inference_engine_t objects. 🤔
   b. Construction and writing of a neural network to a file starting from user-defined weights and biases (useful for unit testing).
   c. Reading a neural network from a file and querying it for basic properties.
   d. Reading and writing from a NetCDF file (for future incorporation into the Inference-Engine library for purposes of reading validation data sets). 🥅

Because the infer procedure is pure, it can be called inside a do concurrent construct, which facilitates concurrent inference using multiple, independent neural networks.

Full Changelog: https://github.com/BerkeleyLab/inference-engine/commits/0.1.0