Normalize long trailing decimal zero suffixes

[perfloop-agent]

Summary

Long ambiguous decimals can reach the bigint comparison fallback. That path materialized trailing decimal zeros even though they add no information.

This change:

• checks for 16 terminal zeros before it scans the full suffix;
• trims only logical integer and fraction spans;
• keeps the original parse metadata and scientific exponent;
• keeps the rare reparse and full trim out of line;
• leaves inputs below the 16-zero cutoff on the existing path.

All-zero coefficients remain unchanged. Integer zeros are trimmed only after an all-zero fraction. Public result bits, consumed pointers, and error codes stay identical.

Performance

All measurements use the public fast_float::from_chars<double> path.

workload	result
0-15 terminal zeros	no regression; 1192.1 ns to 1153.1 ns in the accepted run
16-63 terminal zeros	6.4% faster
64-4096 terminal zeros	48.6% faster
ordinary decimals	no regression; 12.51 ns to 11.73 ns

A later repeat of the exact zero-bound 0-15 measurement was statistically undecidable because its interval crossed zero. It did not show a material regression. The production headers are unchanged from the accepted result, and an independent cross-build kept the public-path gains when benchmark layout was held constant.

Testing

• Full CMake suite: 15/15.
• C++20 focused tests and sanitizer tests pass.
• AppleClang C++11, C++17, and C++20 smoke builds pass.
• Differential public parsing matches for 60,185 char inputs, 6,208 wide-character inputs, and 25,002 float/double inputs.
• The opt-in benchmark keeps only public from_chars modes and route verification. It has no direct internal-target benchmark.

Full verification record: https://app.perfloop.ai/t/oss/cases/case_xwm9h4yejp