Marshal benchmark is outdated

[funny-falcon]

Commit 7581017 unavoidably increases allocations for buffer. But benchmark results in README were not updated to reflect that changes.

Why those commit were done at all? Why allocation care core were replaced with allocation unaware?

I see that simplifies code a lot. But couple of tweaks could make new simple code also allocation aware without introducing much complexity.

[funny-falcon]

Though, if stream is used without passing io.Writer to, then new code certainly works better.