Format decimal numbers with BigDecimal instead of Float

[gaoflow]

Decimal#apply coerces every number to Float before formatting, so the shared Number engine (behind Decimal, Currency and Percent) inherits three defects from that one conversion:

• Half-up rounding. round_to uses Float#round; CLDR / UTS Export exemplar characters #35 rounds half-even, as do ICU and Babel. Decimal.new("0").apply(2.5) returns "3" (should be "2"), and apply(0.125) at "0.00" returns "0.13" (should be "0.12").
• Scientific notation. Float#to_s renders 10**15 as "1.0e+15" and split(".") shreds it, so Number.new("#,##0").apply(10**15) returns "1.0e+15" -- invalid output for an exact value. The repo's own Integer formatter already returns "1,000,000,000,000,000" here.
• Precision loss. Float(12345678901234567890) drops digits before formatting.

The fix routes formatting through BigDecimal and rounds half-even, so rounding is exact and large integers never reach scientific notation. The rounding change touches only exact binary halves (n.5, n/8); artifacts such as 2.675 (not exactly 2.675 in binary) are unchanged. Expected values in the tests were cross-checked against ICU and Babel 2.17.

Tested: the decimal, number, currency and percent suites pass (e.g. bundle exec ruby test/format/decimal/number_test.rb). The currency and percent test files were empty and now cover the shared rounding path.