round_float with TiesToAway rounds inputs just below half the smallest subnormal up #67

Hi! I was comparing gfloat's P3109 rounding with FloatLib, a Lean library for these formats, and found a small difference between round_float and round_ndarray.

For K = 8, P = 3, signed, extended, the smallest positive value is 2^-17. An input just below half of that should round to 0 under TiesToAway. round_ndarray returns 0, but round_float returns 2^-17:

import math
import numpy as np
import gfloat
from gfloat.formats import format_info_p3109

fi = format_info_p3109(8, 3, gfloat.Signedness.Signed, gfloat.Domain.Extended)
x = math.nextafter(2.0**-18, 0)  # 0x1.fffffffffffffp-19

print(gfloat.round_float(fi, x, gfloat.RoundMode.TiesToAway, False))
# Actual: 7.62939453125e-06 (2^-17)
# Expected: 0.0
print(gfloat.round_ndarray(fi, np.array([x]), gfloat.RoundMode.TiesToAway, False))
# [0.]

Reproduced on main at caea0b1a, with Python 3.12.13 and NumPy 2.5.4.

Here delta is 0.5 - 2^-54, and delta + 0.5 rounds to 1.0 in binary64. The fix in src/gfloat/round.py appears to be the comparison round_ndarray.py already uses:

             case RoundMode.TiesToAway:
-                should_round_away = delta + 0.5 >= 1.0
+                should_round_away = delta >= 0.5

With this change round_float and round_ndarray agree on inputs below, at and above the midpoint, for both signs. I'm happy to send a PR with a regression test :)