Restore massive RAMBO reshuffling in CUDACPP SA RAMBO path

[Copilot]

rambo.h in the CUDACPP SA plugin templates was missing the massive-particle reshuffling step, producing incorrect phase-space kinematics/weights for processes with non-zero final-state masses. This change restores the MG5_aMC@NLO-aligned massive branch and wires final-state masses through the sampling stack.

• RAMBO algorithm: restore massive branch (rambo.h)

  • Reintroduced post-massless Newton-Raphson reshuffling (x solve) for non-zero final-state masses.
  • Applied momentum rescaling and energy recomputation for outgoing particles.
  • Added massive weight correction term (wtm) and warning counters for massive-path overflow/underflow.
  • Extended ramboGetMomentaFinal(...) signature with const fptype* massesFinal (defaulting to null for massless behavior).

• Sampling kernel interface: carry mass data end-to-end

  • Extended SamplingKernelBase, RamboSamplingKernelHost, and RamboSamplingKernelDevice to accept/store final-state masses.
  • Host path now calls RAMBO with massesFinal.
  • Device path now allocates/copies/frees a device-side final-mass buffer and passes it to the kernel launcher.

• Process integration: source final-state masses from process model

  • Enabled CPPProcess::getMasses() accessor in template process class.
  • In check_sa.cc, extracted final-state slice via process.getMasses().data() + CPPProcess::npari and passed it into RAMBO sampling kernel construction.

// rambo.h
ramboGetMomentaFinal( energy, rndmom, momenta, wgts, massesFinal );

// check_sa.cc
const fptype* massesFinal = process.getMasses().data() + CPPProcess::npari;
prsk.reset( new RamboSamplingKernelHost( energy, hstRndmom, hstMomenta, hstWeights, nevt, massesFinal ) );