This should lower memory consumption, and fixes OOM in some scenarios.

Before, buggy behavior:

$ yes | tr -ds a b | head -n5  # GNU
y
y
y
y
y
$ yes | cargo run -- -ds a b | head -n5  # WARNING: THIS CAUSES OOM!
^C
[$? = 130]
$ yes yes | tr -s ye s | head -n5  # GNU
s
s
s
s
s
$ yes yes | cargo run -- -s ye s | head -n5  # WARNING: THIS CAUSES OOM!
^C
[$? = 130]

After:

$ yes | cargo run -- -ds a b | head -n5
y
y
y
y
y
$ yes yes | cargo run -- -s ye s | head -n5
s
s
s
s
s
$

It might be tempting to factor out the common line translate_input(&mut locked_stdin, &mut buffered_stdout, op); from all if-branches. However, note that op has distinct types in each case, and virtualizing it (e.g. let op: Box<dyn SymbolTranslator>) would incur additional runtime costs. That cost seems unnecessary just to gain a little bit of beauty.

I'm not really sure how to test this. Ideally we could pipe in some data, and NOT close the stdin pipe, and see whether the program produces any partial output. However, I didn't see any methods/functions for doing so, and it sounds like it's too hard to get right to be worth it.

Maybe I could write a BENCHMARKING.md file, like for shuf? What do you suggest?