Optimize NTT calls

[guy-ingo]

Currently, the main flow uses ntt_end2end which generates and copies the twiddle factors to the gpu on every run.

This can be optimized by:

1. generating twiddle factors for the maximal size needed for the flow (which is 4096) at the beginning of the protocol, and copy them to the device.
2. Change the main flow to use ntt and ecntt instead of the end2end functions.
3. delete the twiddle factors from the device.