Rust: Zero-copy & zero-allocation saves a lot

NOTES.md

Some benchmark results, for:

1. Setting up a MemReader.
2. Setting up a MemReader and reading 130 tokens as sentences (with lots of String and Vector).
3. Reading 1 token, with no copies or allocations.

test conll::memreader_overhead         ... bench:       195 ns/iter (+/- 22) = 22394 MB/s
test conll::sentence_reader_iter_bench ... bench:    269884 ns/iter (+/- 40269) = 16 MB/s
test conll::token_from_str_bench       ... bench:        99 ns/iter (+/- 24) = 313 MB/s

We can see that the MemReader incurs negligible overhead, and Token::from_string is fast enough that it won't be a bottleneck (apparently the Snappy decompression algorithm, one of the fastest in common use, runs at around 500MB/sec on a single i7 core).

Next up: Modifying the sentence reader to require zero amortized allocations.