Update layers and losses for mixed precision compatibility.

[willgraf]

What

• Change references to K.floatx to refer to inputs.dtype or compute_type.
• Cast y_true to y_pred.dtype to preserve dtype in losses.py.

Why

• Setting floatx to float16 is not recommended for training, and instead the global precision policy should be changed to mixed_float16. However, the Location and Normalization layers were set to output K.floatx() which conflicts with the global precision policy. Instead, setting the dtype to the internal compute_type or inputs.dtype preserves the mixed dtype policy.
• Fixes Enable training/inference with half-precision (float16) #112