Improve attention performance for qwen2.5 & deepseek

[orionpapadakis]

This PR adds a new Tornado kernel for parallel attention and improves grid sizes calculations.

[mikepapadim]

put it into a seperate method:
and make 64 constant etc

    public static int computeOptimalLocalSize(int headSize) {
        int optimalLocalSize = Math.min(headSize, 64);

        if (headSize % optimalLocalSize != 0) {
            for (int size = optimalLocalSize; size >= 1; size--) {
                if (headSize % size == 0) {
                    return size;
                }
            }
        }

        return optimalLocalSize;
    }