Fix desi_proc specex wrapper MPI rank bug for single camera

[marcelo-alvarez]

Fixes #1528 by ensuring that 21 MPI processes are used for a single camera.

Following tests with this branch pass:

Cameras: z1
  haswell: srun -N 1 -n 21 -c 3
      knl: srun -N 1 -n 21 -c 12
Cameras: b1
  haswell: srun -N 1 -n 21 -c 3
      knl: srun -N 1 -n 21 -c 12
Cameras: r135
  haswell: srun -N 1 -n 31 -c 2
      knl: srun -N 1 -n 31 -c 8
Cameras: a3
  haswell: srun -N 1 -n 31 -c 2
      knl: srun -N 1 -n 31 -c 8
Cameras: a0123456789
  haswell: srun -N 10 -n 301 -c 2
      knl: srun -N 10 -n 301 -c 9

Please review and merge if appropriate.

[coveralls]

Coverage decreased (-0.008%) to 25.646% when pulling 8e41121 on mpirankbug into 3adabaf on master.

[sbailey]

Thanks @marcelo-alvarez . Please check running a case with an odd number of cameras, where it selects the number of ranks as 10*n+1, e.g. the r135 case above. It seems to successfully run specex and generate the initial fit-psf-??-????????.fits files within a few minutes, but then it gets wedged before it gets to the desi_proc logic for interpolating over bad fibers and the jobs just keeps running without apparently doing anything. Running with an even number of cameras such that the selected number of ranks is 20*n+1 seems to work fine.

[marcelo-alvarez]

Apologies for not verifying the previous solution actually runs without crashing. The latest commit runs without crashing for r135 and is expected to do so for all possible camera combinations, from a single camera to all 30.

[sbailey]

Looks good. Thanks for including the comment "lowest multiple of 20 exceeding 10 per camera" explaining the magic 20 math for our future selves. Merging now.