Normalize long trailing decimal zero suffixes

[perfloop-agent]

Summary

When a long ambiguous decimal falls back to big-integer comparison, digit_comp materializes every digit into the bigint — including an arbitrarily long suffix of trailing zeros, which inflates limb construction and the subsequent comparison for no informational gain.

This PR trims a long trailing-zero suffix before materializing the coefficient:

• has_trailing_zero is an O(1) last-character check, so inputs without a trailing zero pay a single branch.
• trim_zeros_from_end scans backward (word-at-a-time, then per-character); suffixes shorter than 16 code units are kept as-is (reverse-scan setup doesn't pay for itself below that).
• Trimming K zeros shortens the digit count by K, which raises the derived exponent (sci_exp + 1 - digits) by exactly K — the represented value mantissa × 10^exp is unchanged, so rounding decisions are identical.
• Integer-part zeros are only trimmed when the entire fraction was zeros (e.g. 120.3000 keeps the zero in 120), and an all-zero coefficient is never reduced to empty spans.

The scientific exponent is computed from the original parsed number before trimming; only the local spans passed to parse_mantissa change.

Performance

Measured with the opt-in benchmark added in this PR (-DFASTFLOAT_TRAILING_ZERO_BENCHMARK=ON), over ambiguous 20-, 30-, and 120-digit cores with 64–4096 trailing zeros:

metric	before	after
direct digit_comp (per conversion)	~1445 ns	~200 ns
public from_chars<double> (per parse)	~2789 ns	~1526 ns

Independently reproduced on Apple Silicon (clang, arm64): ~1035 ns → ~114 ns per digit comparison on the same shapes.

Testing

• Full CMake test suite passes across C++11/C++17/C++20 modes, including the constexpr basictest and a sanitizer run of the new trailing-zeros test.
• New trailing_zeros_test covers the 16-unit cutoff boundary, integer-vs-fraction trimming (120.3000-style), all-zero coefficients, and wide-character variants.
• Deterministic differential probes over several hundred thousand long/malformed/trailing/boundary inputs: double, float, and char16_t/char32_t/wchar_t outputs (bits, end pointer, and error code) are identical before and after.

Note: this touches the same slow-path file as #402; whichever lands second will need a trivial test-file rebase.

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