feat: Implement BufWriter optimization for du stdout output (fixes #9146)

[naoNao89]

Replace direct print! statements with 64KB buffered writes to reduce syscall overhead for du -a on large directories. This addresses the performance issue where each file entry triggers multiple stdout writes.

Changes:

• Add BufWriter field to StatPrinter struct with 64KB buffer
• Convert print_stat method to use buffered writes instead of direct output
• Add Drop implementation to ensure proper buffer flushing
• Add comprehensive benchmarks to validate performance improvement

Performance: Reduces syscalls from ~16,500 to ~3-5 for 5,500 file directories, providing significant improvement for the issue #9146 use case.

[codspeed-hq]

CodSpeed Performance Report

Merging #9163 will improve performances by 12.27%

_{Comparing naoNao89:feature/du-buffer-optimization-9146 (2887ac4) with main (0058316)}

Summary

⚡ 4 improvements
✅ 121 untouched
🆕 2 new

Benchmarks breakdown

	Benchmark	BASE	HEAD	Change
⚡	du_all_balanced_tree[(4, 3, 10)]	1.6 ms	1.5 ms	+7.44%
🆕	du_all_extreme_wide_tree[(2000, 500)]	N/A	10.3 ms	N/A
🆕	du_all_stress_balanced_tree[(3, 5, 6)]	N/A	1.1 ms	N/A
⚡	du_all_wide_tree[(5000, 500)]	18.1 ms	16.2 ms	+12.27%
⚡	du_balanced_tree[(5, 4, 10)]	9.3 ms	8.8 ms	+4.82%
⚡	du_human_balanced_tree[(5, 4, 10)]	10.4 ms	10.1 ms	+3.38%

[github-actions]

GNU testsuite comparison:

GNU test failed: tests/du/inodes. tests/du/inodes is passing on 'main'. Maybe you have to rebase?
Skip an intermittent issue tests/tail/overlay-headers (fails in this run but passes in the 'main' branch)

[Its-Just-Nans]

Hello

In your code you are using a capacity of 65536

So it seems to be 64 KiB and not 64KB

[naoNao89]

@Its-Just-Nans you're right, now am waiting for Github to restore my fork :)))

[Its-Just-Nans]

hi

others questions

• Why not putting the BufWriter::with_capacity(65536, stdout()) inside the function print_stats
• Why not locking the stdout ? BufWriter::with_capacity(65536, stdout().lock())

[naoNao89]

the current BufWriter setup in the struct is cleaner, and stdout().lock() isn't needed since du's output is already serialized. The commit in goal was just buffering for syscall reduction, not changing locking.