move away from GSL functions.

[stevengj]

Now defaults to Mersenne Twister random-number generation (included under 3-clause BSD license) and defaulting to POSIX Bessel functions rather than the much-slower GSL ones for near2far calculation (for #861).

[stevengj]

@oskooi, try the near2far calculation with this PR — it should be much faster (in 2d) with the new Bessel functions.

[oskooi]

The following are the wall-clock Field time usage statistics for examples/binary_grating_n2f.py using resolution=25, nfreq=50, and nperiods=40 with MPI (np=2) comparing master and this PR:

master

unit cell with periodic boundaries

Field time usage:
        connecting chunks: 0.00340685 s
            time stepping: 1.70402 s
            communicating: 0.368193 s
     Fourier transforming: 0.0024059 s
        getting farfields: 1525.26 s
          everything else: 2.16774 s

finite periodic grating with flat surface termination extending into PML

Field time usage:
	connecting chunks: 0.179651 s
            time stepping: 270.982 s
            communicating: 9.90172 s
     Fourier transforming: 125.645 s
	getting farfields: 763.141 s
          everything else: 2.1376 s

this PR

unit cell with periodic boundaries

Field time usage:
	connecting chunks: 0.00318903 s
            time stepping: 1.71046 s
            communicating: 0.369378 s
     Fourier transforming: 0.00197067 s
        getting farfields: 843.431 s
          everything else: 2.1393 s

finite periodic grating with flat surface termination extending into PML

Field time usage:
	connecting chunks: 0.187708 s
            time stepping: 272.32 s
            communicating: 9.89654 s
     Fourier transforming: 126.648 s
	getting farfields: 420.911 s
          everything else: 2.61026 s

The speedup in get_farfields is significant as the runtime is nearly halved for both the unit- and super-cell calculations: 1525.26 s vs. 843.431 s and 763.141 s vs. 420.911 s.