Add mixed precision solve capabilities

[pratikvn]

Add capabilities to solve with mixed precision. In particular, gather and scatter operations can additionally perform averaging, sum, difference operations before copying.

Additionally, some minor improvements are also made:

• Remove the dependency on boost (Add the get_mpi_type function internallly).
• Fix the benchmarking documentation.
• Use executor copies instead of CudaMemcpy's
• Move gather and scatter to separate classes and namespace qualify them.
• Add timestamping and local residual printing directly (computed in the convergence checks) instead of from loggers.

[pratikvn]

It has been verified that this gives similar performance to the develop branch before merging this PR.