Fix directional rounding (ROUND_CEILING/ROUND_FLOOR) for negative numbers

[vineethsaivs]

Overview

Fixes #1096.

ROUND_CEILING and ROUND_FLOOR produced the wrong result when formatting a
negative number:

from babel.numbers import decimal, format_decimal

with decimal.localcontext(decimal.Context(rounding=decimal.ROUND_CEILING)):
    print(format_decimal(decimal.Decimal("-100.75"), format="#", locale="en_US"))
# before: -101
# after:  -100   (ceil(-100.75) == -100)

Root cause

NumberPattern.apply() separates the sign from the value up front
(value = abs(value).normalize()), formats the magnitude, and renders the sign
separately through the pattern's affixes. Because the sign was already gone by
the time rounding happened, the directional rounding modes (ROUND_CEILING,
ROUND_FLOOR) were applied to the magnitude and therefore rounded toward the
wrong infinity for negative numbers. Effectively ROUND_CEILING behaved like
ROUND_UP and ROUND_FLOOR like ROUND_DOWN for negatives.

This affected every rounding site that operates on the magnitude:
_quantize_value (plain and fractional patterns, e.g. # and #,##0.00, plus
the scientific-notation mantissa) and _format_significant (significant-digits
patterns such as @@).

Fix

Round using the value's original sign in both sites, then format the magnitude
(the sign is still rendered via the affixes, so output formatting is otherwise
unchanged). The sign-symmetric modes (ROUND_HALF_*, ROUND_UP, ROUND_DOWN,
ROUND_05UP) and all positive-number formatting are unaffected; verified
against decimal.Decimal.quantize on the signed value across every rounding
mode.

Tests

Added test_format_decimal_negative_directional_rounding (parametrized over the
rounding modes) and test_format_decimal_negative_directional_rounding_significant.
Both fail on master for the directional modes and pass with this change. The
full tests/test_numbers.py suite passes (137 tests).