[FTRL paper](https://research.google.com/pubs/archive/41159.pdf).
*   `Optimizer.__init__` now accepts a `gradient_aggregator` to allow for
    customization of how gradients are aggregated across devices, as well as
    `gradients_transformers` to allow for custom gradient transformations (such
    as gradient clipping).
*   Improvements to Keras preprocessing layers:

BLK_OUT R4 #define BLK_KEY R5 #define KEY_LEN R6 #define BLK_IDX R7 #define IDX R8 #define IN_LEN R9 #define COUNTER R10 #define CONPTR R14 #define MASK V5 // Implementation of the counterCrypt function in assembler. // Original loop is unrolled to allow for multiple encryption // streams to be done in parallel, which is achieved by interleaving // vcipher instructions from each stream. This is also referred to as // stitching, and provides significant performance improvements. // Some macros are defined...