of simply using the `dtype` & `shape`.
    *   [Improve support](https://github.com/tensorflow/tensorflow/issues/33862)
        for dynamically-sized TensorArray inside `tf.function`.
*   `tf.math`:
    *   Narrow down `argmin`/`argmax` contract to always return the smallest
        index for ties.
    *   `tf.math.reduce_variance` and `tf.math.reduce_std` return correct

h.inner.Write(h.ipad) // If the underlying hash is marshalable, we can save some time by saving a // copy of the hash state now, and restoring it on future calls to Reset and // Sum instead of writing ipad/opad every time. // // We do this on Reset to avoid slowing down the common single-use case. // // This is allowed by FIPS 198-1, Section 6: "Conceptually, the intermediate // results of the compression function on the B-byte blocks (K0 ⊕ ipad) and // (K0 ⊕ opad) can be precomputed once, at the time of generation...