libc: prevent incorrect %a/%La rounding at full precision

In hdtoa() and hldtoa(), rounding is incorrectly applied when
the requested precision exactly matches the number of significant
hexadecimal digits. In this case, the redux adjustment can trigger
an unintended exponent increment and shift the rounding position
left by one bit. This causes the least significant digit to be
rounded incorrectly.

The fix adds a new condition based on MAX_HEX_DIGITS (derived from
MANT_DIG) so that rounding is performed only when precision is
strictly less than the number of significant digits. This avoids
the unintended shift while preserving correct rounding for other
cases.

A new regression test
(printfloat_test:hexadecimal_rounding_offset_eq_exp) covers both
the binary64 (%.13a) and binary128 (%.28La on arm64) cases that
previously fail, ensuring the bug does not regress.

Note: MAX_HEX_DIGITS represents the maximum number of hexadecimal
digits needed to express the mantissa. It is computed by subtracting
the implicit integer bit from [L]DBL_MANT_DIG, dividing the remaining
mantissa bits by 4 (with +3 to round up any remainder), and finally
adding +1 for the leading integer digit. This makes its meaning
explicit and distinct from SIGFIGS, which serves a different purpose.

Fixes: 76303a9735ee ("Make several changes to the way printf handles hex floating point (%a):")
Signed-off-by: Osamu Sho <osamusho@gmail.com>
Reviewed by: imp,jlduran
Pull Request: https://github.com/freebsd/freebsd-src/pull/1837