perf: Speedup ndjson reader ~40%

[ChayimFriedman2]

This speeds it up by 40% in the following benchmark:

use criterion::{criterion_group, criterion_main, Criterion};
use mimalloc::MiMalloc;
use polars::prelude::*;
use polars_core::POOL;
use rand::{thread_rng, Rng};

#[global_allocator]
static MIMALLOC: MiMalloc = MiMalloc;

fn from_json(json: &[u8]) -> DataFrame {
    JsonReader::new(std::io::Cursor::new(json))
        .with_json_format(JsonFormat::JsonLines)
        .set_rechunk(false)
        .finish()
        .unwrap()
}

fn my_benchmark(_c: &mut Criterion) {
    POOL.install(|| {
        let mut c = Criterion::default().configure_from_args();

        const SIZE: i32 = 5_000_000;
        let mut rng = thread_rng();
        let mut df = df![
            "a" => (0..SIZE).map(|_| rng.gen::<i32>()).collect::<Vec<_>>(),
            "b" => (0..SIZE).map(|v| v.to_string()).collect::<Vec<_>>(),
        ]
        .unwrap();

        let mut json = Vec::new();
        JsonWriter::new(&mut json)
            .with_json_format(JsonFormat::JsonLines)
            .finish(&mut df)
            .unwrap();

        c.bench_function("JSON Lines Deserialization", |b| {
            b.iter(|| from_json(&json))
        });
    });
}

criterion_group!(benches, my_benchmark);
criterion_main!(benches);

I had a plan to improve it more, but I can't find time for that and this will involve bigger changes, even a rewrite of the mechanism, while the changes in this PR are simple and effective, so I thought I'll just send them.

Best reviewed commit-by-commit.

A warning from the second commit, repeated here for noticeability:

This could break people's code since we will not split correctly (and thus error) if one object spans two lines or two objects are in the same line. However, such code was already broken, since NDJSON is not allowed to contain any line breaks. If this is a concern, it is possible (at some perf degradation) to check for }\n instead of \n alone, and that will make this basically equivalent to the splitting logic we have for threads.

[codecov]

Codecov Report

Attention: Patch coverage is 96.96970% with 1 line in your changes missing coverage. Please review.

Project coverage is 80.31%. Comparing base (aa1950c) to head (ecf762d).
Report is 106 commits behind head on main.

Files	Patch %	Lines
crates/polars-io/src/ndjson/core.rs	96.96%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main   #18197      +/-   ##
==========================================
- Coverage   80.35%   80.31%   -0.05%     
==========================================
  Files        1492     1498       +6     
  Lines      196332   198748    +2416     
  Branches     2813     2833      +20     
==========================================
+ Hits       157759   159618    +1859     
- Misses      38052    38603     +551     
- Partials      521      527       +6     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[ritchie46]

Really nice PR and great review with those commit messages. I have one question.