Fix regtests with Real BCs and correct a bug

[AMLattanzi]

The previous PR that corrected issues with numerical diffusion provided some conflicts with cases that use Real BCs. The end of the fast integrator will call apply_bcs which will fill the fast conserved vars. It will first fill rho with std::max(ng_cons,ng_vel+1) ghost cells. So this operation will occur over all the rho ghost cells. The second call to FillIntermediatePatch but uses ng_cons directly. By using ng_cons=ng_cons_tot-1 we have an overhead savings since rhoTheta is not needed in the last ghost cell since it is not used to convert VelocityToMomentum.

However, numerical diffusion still did not work with the Real BC case after correcting the ng_cons size. The root cause of this was found to be a bug in FillPatchIntermediate where the fill_from_realbdy call was not aware of the starting component icomp_cons. Since this value was hard-coded to 0 we always filled rho rather than filling rho in the first FillPatchIntermediate call and then rhoTheta in the second. Numerical diffusion now works after this correction. A remake of the benchmarks is expected.