Allowed saving model with different dtype (float16, float32)

[pure-rgb]

Whether enable mixed precision with dtype float16 or not (float32), it would be better to have option to save model with different dtype.

keras.callbacks.ModelCheckpoint(
    filepath,
    monitor="val_loss",
    verbose=0,
    save_best_only=False,
    save_weights_only=False,
    mode="auto",
    save_freq="epoch",
    initial_value_threshold=None,
    dtype='float16'  # < ---------------------- NEW ADD
)

• It is usefull for some cases where the mode is too big to use. Saving it in float16 format, would definitly help.
• The savaed float16 model should work properly later if mixed precision is not enabled, means float32.
• If mixed precision is enabled (float16), the ModelCheckpoint should be independent, means if dtype=float32 is set for this callback, it should work.