updated the n_task calculation to use max number of threads possible. This has improved the prompt eval performance by around 5% for DOT kernels and by around 10% for MMLA kernels on AWS Graviton3.
