Fix si_format() producing out-of-range mantissa like "1000.0 k"

[patchwright]

Description

si_format() can print a mantissa of 1000 with the smaller prefix, which is
outside the [1, 1000) range each SI prefix represents:

>>> from si_prefix import si_format
>>> si_format(999999)
'1000.0 k'         # expected: '1.0 M'
>>> si_format(999999, precision=0)
'1000 k'           # expected: '1 M'
>>> si_format(999.9, precision=0)
'1000 '            # expected: '1 k'
>>> si_format(999999999)
'1000.0 M'         # expected: '1.0 G'

Root cause

split() selects the prefix by promoting the mantissa while value >= 1000.0,
but it works on the unrounded value. si_format() then rounds the mantissa to
precision digits ("%.1f" % 999.999 → "1000.0"). So a value just below 1000
rounds up to 1000 after the prefix has already been chosen.

Fix

Make split()'s promotion test the value rounded to precision, so a mantissa
that will round to 1000 is carried into the next prefix:

if round(value, precision) >= 1000.0:
    value /= 1000.0
    expof10 += 3

Values below the rounding boundary (si_format(999500, 1) == "999.5 k") and
higher-precision inputs (si_format(999999, 4) == "999.9990 k") are unaffected.

Tests

• Added test_si_format_rounding_rollover with the cases above; it fails on
  master and passes with the fix.
• The existing 38 si_format cases are unchanged — full suite (112 tests) green.