Implicit Field Solve Preconditioner based on Curl-Curl Operator

[debog]

Implemented a preconditioner for the implicit E-field solve using the AMReX curl-curl operator and the MLMG solver.

• Introduced a Preconditioner base class that defines the action of a preconditioner for the JFNK algorithm.
• Implemented the CurlCurlMLMGPC that uses the multigrid solution for the curl-curl operator (implemented in AMReX) to precondition the E-field JFNK solve.

Other changes needed for this:

• Partially implemented a mapping between WarpX field boundary types and AMReX's linear operator boundary types.
• Added some functionalities to ImplicitSolver class that allows preconditioners to access WarpX info (like Geometry, boundaries, etc).

Some premilinary wall times for:

Test: inputs_vandb_2d
  Grid: 160 X 160
  dt: 0.125/wpe = 2.22e-18 (dt_CFL = 7.84e-19 s, CFL = 2.83)
  Time iterations: 20

Solver parameters:
  newton.max_iterations = 10
  newton.relative_tolerance = 1.0e-12
  newton.absolute_tolerance = 0.0
  gmres.max_iterations = 1000
  gmres.relative_tolerance = 1.0e-8
  gmres.absolute_tolerance = 0.0

Avg GMRES iterations: ~3 (wPC), ~27 (noPC)

with 32^2 particles per cell:

Lassen (MPI + CUDA)
-------------------
  Box  GPU   Walltime (s)
             wPC       noPC
   1    1    2324.7    15004.1
   4    1    2306.8    14356.8
   4    4     758.9     3647.3

Dane (MPI + OMP)
----------------
  Box  CPU  Threads   Walltime (s)
                      wPC      noPC
   1    1      1      6709.3   43200.0*
   1    1      2      3279.1   22296.1
   1    1      4      1696.3   11613.2
   1    1      8      1085.0    6911.4
   1    1     16       724.3    4729.0
   4    1      1      5525.9   33288.8
  16    1      1      4419.4   28467.8
   4    4      1      1324.4    9121.1
  16   16      1       524.9    3658.8

* 43200.0 seconds is 12 hours (max job duration on Dane);
the simulation was almost done (started the 20th step).

with 10^2 particles per cell:

Lassen (MPI + CUDA)
-------------------
  Box  GPU   Walltime (s)
             wPC       noPC
   1    1    365.0     1443.5 
   4    1    254.1      927.8 
   4    4    133.1      301.5 

Dane (MPI + OMP)
----------------
  Box  CPU  Threads   Walltime (s)
                      wPC      noPC
   1    1      1      440.8    2360.5     
   1    1      2      241.7    1175.8 
   1    1      4      129.3     727.0 
   1    1      8       94.2     407.5 
   1    1     16       74.3     245.6 
   4    1      1      393.3    1932.5 
  16    1      1      337.6    1618.7 
   4    4      1       92.2     479.1 
  16   16      1       58.1     192.6 

[JustinRayAngus]

Making this a member of the base class is fine for now. In a future PR, we can remove it by passing in m_theta*a_dt to the solver. This will require modifications to several files in the ImplicitSolver folder and in the NonlinearSolvers folder, but it will make the code more tidy. I can do this after this PR is merged.

[JustinRayAngus]

This PR looks good to me.