fix: preserve zero fractions when formatting integral floats

[rksharma-owg]

std.format("%f", 1e20) produces "100000000000000000000.729344", even though the input is an exactly representable whole number. Multiplying that value by the requested precision introduces rounding error before the remainder is calculated.

Preserve the whole value and a zero fraction when the input is integral. Non-integral inputs retain the existing rounding path. Regression tests cover large positive and negative integers, the largest safe integer, precision zero, signs, padding, and a fractional value that rounds into the next integer.

Refs #1331. This addresses its spurious %f fractional digits; the separate exact-decimal %d conversion discrepancy remains outside this change. This is independent of #1332, which validates dynamic width and precision.

Validation: the new regression failed before the fix; all 173 scripts in test_suite/run_tests.sh pass after rebuilding. git diff --check and formatting of test_suite/format.jsonnet pass. The stdlib formatter reports only existing differences outside the changed lines.