Improve performance

This isn't very high priority as serialization/deserialization isn't really a bottleneck in Substrate AFAIK, although the current implementation leaves some performance on the table.

For reference, I've hooked up parity-scale-codec to the rust_serialization_benchmark benchmarks and here are the results compared to speedy which is my personal serialialization/deserialization crate that I've optimized for speed and with almost exactly the same serialization format as parity-scale-codec (it doesn't special-case Option<bool> and has a different varint format, but the rest should serialize the same, sans defaults). Here are the results on my machine:

Start by reviewing the referenced rust_serialization_benchmark results and the parity-scale-codec implementation, then compare the reported cases with speedy. The issue does not name repository files, tests, or a target improvement; completion would require identifying a concrete optimization and validating it against the referenced benchmarks.