Speedup acquired:

Before opt. (the script runs pyfaust.FGFT.eigtj() on a 256 community graph Laplacian asking for J=1056 and t=128):
$ python3 test_fgft_time_givens_par.py 1000
total time (1000 runs): 102.907594
mean time: 0.102907594

After opt.
$ python3 test_fgft_time_givens_par.py 1000
total time (1000 runs): 73.510564
mean time: 0.073510564

GivensFGFT::update_L() was cut in two parts update_L_first() and update_L_second(), this first one computing S'*L, the second L*S (with L = S'*L here). The goal was to reuse for GivensFGFTParallel the optimization made for GivensFGFT by c95de0c5 (without duplicating the code computing the linear combinations of rows or columns of L).