Method parallel_for

csrc/adam/cpu_adam_impl.cpp:406–419

Split [0, total) into one chunk per thread. Chunk boundaries are rounded up to a multiple of `align` so each slice's AVX/scalar split lines up with the whole-tensor kernel's split -- otherwise an element could be computed by AVX (FMA) in one layout and the scalar tail (mul+add) in another, which differ in the last bit.

    // kernel's split -- otherwise an element could be computed by AVX (FMA) in one layout
    // and the scalar tail (mul+add) in another, which differ in the last bit.
    void parallel_for(size_t total, size_t align, std::function<void(size_t, size_t)> fn)
    {
        {
            std::unique_lock<std::mutex> lk(m_);
            fn_ = std::move(fn);
            total_ = total;
            align_ = std::max<size_t>(1, align);
            done_count_ = 0;
            ++gen_;
        }
        cv_start_.notify_all();
        std::unique_lock<std::mutex> lk(m_);
        cv_done_.wait(lk, [this] { return done_count_ == n_; });
    }

private:
    void worker(size_t tid, int core)