Performance tweaks

Currently there are performance issues due to indirection between decoding a marker and a value itself; and non-inlining in the core RMP crate. Need to write benchmarks and fix them. The one approach is to mark some functions as `[inline]` or `[inline(always)]`. This works, but I feel it a little bit hacky. The other one - is to match for marker bytes directly and avoid using `read_marker`, but we lose an abstraction then.