Disable stochastic behavior when in inference mode

[jecorona97]

Currently, quantizers peform stochastic operations even in inference mode (K.learning_phase() = 0). This mechanism prevents this and uses the non-stochastic version of the quantizer if in inference mode.

Added

• base_quantizer class with is_training attribute and setter method
• unit tests for stochastic quantizers inference mode behavior

Modified:

• stochastic_{ternary, binary} now inherit from their non stochastic counterparts to delegate inference mode computaiton
• Order of quantizer definition to perfom inheritance
• Used the super invokation method in quantizers instead of referring to the base class explicitly
• set the proper learning_phase in qactivation_test, qconvolutional_test
• Fixed _round_through and _clip_power_of_two and their references by using smart_cond to disable stochastic behavior

Some fixes by Claudionor:

• Fixed multiple instance of tensorflow conditionals in quantizers.py
• Changed stochastic_ternary test to remove the 0 variance cases