*   Add support for `add_metric` in the graph function mode.
    *   Updating cosine similarity loss - removed the negate sign from cosine
        similarity.
    *   Changed default for gradient accumulation for TPU embeddings to true.
    *   Adds summary trace API for collecting graph and profile information.
    *   The `precision_mode` argument to `TrtGraphConverter` is now case
        insensitive.

first nonzero coefficient. i := 255 for j := i; j >= 0; j-- { if aNaf[j] != 0 || bNaf[j] != 0 { break } } multA := &projCached{} multB := &affineCached{} tmp1 := &projP1xP1{} tmp2 := &projP2{} tmp2.Zero() // Move from high to low bits, doubling the accumulator // at each iteration and checking whether there is a nonzero // coefficient to look up a multiple of. for ; i >= 0; i-- { tmp1.Double(tmp2) // Only update v if we have a nonzero coeff to add in. if aNaf[i] > 0 { v.fromP1xP1(tmp1) aTable.SelectInto(multA,...