Implement MPI support in CUDA Quantum

[amccaskey]

Add optional MPI support to CUDA Quantum.

Implement observe() functions that distribute expectation value computations across available GPUs on a multi-node, multi-gpu system.

User surface:

# User must initialize MPI (if using MPI)
cudaq.mpi.initialize()
...
ansatz, parameters = cudaq.make_kernel(list)
... 
hamiltonian = ...
# MPI distribution
energy = cudaq.observe(ansatz, hamiltonian, args, execution=cudaq.par.mpi)
# Single-node, multi-GPU distribution
energy = cudaq.observe(ansatz, hamiltonian, args, execution=cudaq.par.thread)
# No Parallelism
energy = cudaq.observe(ansatz, hamiltonian, args)

# If initialized, must finalize
cudaq.mpi.finalize()

// Must initialize if using MPI
cudaq::mpi::initialize();
...
auto ansatz = [](std::vector<double> parameters) __qpu__ {
  ...
};
auto H = ... ;
// MPI Distribution
auto result = cudaq::observe<cudaq::par::mpi>(ansatz, H, args...);
// Single-node, multi-GPU Distribution
auto result = cudaq::observe<cudaq::par::thread>(ansatz, H, args...);
// No Parallelism
auto result = cudaq::observe(ansatz, H, args...);
...
// Must finalize MPI
cudaq::mpi::finalize();

We also introduce user API for initializing and finalizing MPI, getting the number of ranks, and getting the current rank.