Always allocates 4 x input layer x first hidden layer bytes

[janjongboom]

The first layer of my model is: 186 input nodes -> 150 hidden nodes. uTensor allocates 111,600 bytes for this. This should not be necessary. My weights are already distributed between -127 and 127, and thus I should be able to pass in a uint8_t RAM tensor and just allocate 27,900 bytes instead.

Test case here: https://github.com/janjongboom/utensor-test although I made it 186x40 because otherwise I couldn't run this.