[metif12]

numfmt: print what GNU prints for a whole number, and upper case the SI suffixes

Four defects, all measured against GNU 9.4 and uutils 0.0.17 rather than reasoned
about. Twenty cases, five numbers in each of --to=si, --to=iec and --to=iec-i plus
five bare numbers: 8 of 20 agreed before, 20 of 20 now.

numfmt 1000           1000.0   ->  1000
numfmt 0.5            1        ->  0.5
numfmt -12.25         -13      ->  -12.25
numfmt --to=si 1500   1.5k     ->  1.5K
numfmt 2000000.6      2.0000006e+06  ->  2000000.6

• f64.str() is not how GNU prints a number. It keeps a ".0" on a whole one and
  switches to scientific notation on a large fractional one, so both the whole and
  the fractional case were wrong in opposite directions.

• A number with no suffix and no unit to convert is handed back as it is, which is
  what --to=none means. It was being rounded to an integer first, away from zero, so
  0.5 answered 1 and -12.25 answered -13. Rounding now happens only when there is
  something to convert by, and it goes through the function that honours --round,
  which scale_number was not doing: it always rounded from zero.

• SI suffixes are upper case at GNU. The one that said so in this file was
  power_to_suffix, which nothing called; the list that was in use said lower case.
  power_to_suffix is gone rather than left as the second answer to the same
  question.

Where this still differs, on purpose

GNU rejects a suffix in the input unless --from is given:

numfmt 2K            numfmt: rejecting suffix in input: '2K' (consider using --from)

This port has no --from at all, only --from-unit, so the four test cases that pass
a suffix are left asserting what it does instead - 2K is 2000. That is recorded in
the test as this port's behaviour and not GNU's, rather than quietly agreeing with
it. Adding --from is the fix, and it is a separate piece of work.

Tests

src/numfmt is 1 of 1, the twenty-case comparison against GNU 9.4 is 20 of 20, and
v fmt -verify src/numfmt passes. Five of the expectations in test_to_options
changed, and every one of them is a string read off GNU rather than a value this
implementation produced.