numfmt: `--to-unit` uses unscaled value to pick SI/IEC prefix

[sylvestre]

Summary

numfmt --to=<scale> --to-unit=N VALUE selects the output prefix (K, Ki, M, ...) based on the unscaled input VALUE, instead of the final scaled value VALUE / N. GNU selects the prefix based on the scaled result.

Found by fuzz_numfmt.

Reproduction

$ LC_ALL=C /usr/bin/numfmt --to=iec-i --to-unit=885 100000
113

$ target/debug/numfmt --to=iec-i --to-unit=885 100000
0.1Ki

100000 / 885 = 113, which is below 1 Ki (1024), so no suffix should be applied. uutils apparently looks at 100000, picks Ki, and emits 0.1Ki.

Another example:

$ LC_ALL=C /usr/bin/numfmt --to=iec-i --to-unit=885 1000000
1.2Ki
$ target/debug/numfmt --to=iec-i --to-unit=885 1000000
1.1Ki

Scaled value: 1000000 / 885 ≈ 1130, i.e. 1.2Ki (rounding up). uutils seems to compute the suffix from the unscaled value, giving 1.1Ki.