fix(stdlib): renormalize %e/%g mantissa carry and handle precision 0

[jdymitarai]

Description

This PR fixes two related formatting bugs in stdlib/std.jsonnet for %e and %g conversions (reported in #1327):

1. Mantissa Carry Renormalization:
   When formatting floating-point numbers in scientific form (%e / %E / %g / %G), rounding to prec decimal places can round the mantissa up to 10.0 or higher (for example: "%.1e" % 9.99 yielded "10.0e+00" instead of "1.0e+01", and "%.0e" % 9.5e10 yielded "10e+10" instead of "1e+11").
   When mantissa * 10^prec + 0.5 >= 10 * 10^prec, we now detect the carry, increment the exponent by 1, and divide the mantissa by 10 to maintain standard normalized scientific representation (1 <= mantissa < 10).

2. %g Precision of 0 & Boundary Transition:
   Per Python format specifications (which Jsonnet follows), a precision of 0 for %g is treated as equivalent to a precision of 1. Previously, precision 0 resulted in passing negative precision to render_float_dec, leading to corrupted output such as "%.0g" % 1.0 -> "5e+00".
   In addition, when rounding produces an exponent carry (e.g. "%.1g" % 9.9), the effective exponent is updated to 1, correctly switching to scientific notation "1e+01" instead of decimal overflow.

Fixes #1327.

Verification

• Added automated regression tests to test_suite/format.jsonnet for %e, %E, %g, and %G testing mantissa carries, precision 0, and boundary exponent transitions.
• Verified against Python standard library formatting across comprehensive test cases (test_suite/format.jsonnet returns true).
• C++ test suite passes without regressions.