[perf] split: reuse the per-line buffer in round-robin mode

[haydonryan]

Small improvement today.

Moving the allocation out of the loop, and reuse the vec instead of allocating and freeing it every line. Because it's allocated on every line, it will stretch (reallocate) if the line length exceeds the allocated buffer.

Clanker generated:

Vec Mechanics recap:

• First byte pushed → allocates a capacity-8 buffer.
• Every time len hits capacity, it reallocs to 2× the current capacity (amortized O(1) appends).

So for read_until reusing the buffer: it grows through the doubling sequence only while a single line is being read and only until it fits that line; thereafter clear() keeps whatever
capacity that line reached, and shorter subsequent lines reuse it with zero realloc. The 64-byte figure in my earlier answer was just an illustration of "a capacity value in the growth
series," not a fixed allocation size.

Verification (A/B on this branch, main vs split-round-robin-vec-outside-loop)

Release fat-LTO coreutils binaries, split --number=r/4 on a 500k-line file:

Metric	main	branch	Δ
Binary size	13,163,224 B	13,163,416 B	+192 B
Runtime (mean ± σ)	448.1 ± 21.9 ms	396.0 ± 37.1 ms	1.13x faster (±0.12)
User time	73.1 ms	62.6 ms	−10.5 ms
Output	—	—	byte-identical

cargo test -p uu_split passes.

[sylvestre]

could you please add a bench for the r/N mode in benches/split_bench.rs? otherwise codspeed won't track this gain