Preserve sub-second precision when coercing a DateTime in Time.at - #58085

Summary

Time.at accepts an ActiveSupport::TimeWithZone or a DateTime as a single argument (ActiveSupport layers this on top of Time.at). The TimeWithZone branch converts via an exact rational timestamp (to_r), but the DateTime branch converted via a Float (to_f). A Float cannot represent every microsecond exactly, so the resulting Time#usec was wrong for roughly half of all sub-second DateTime values.

Before / After

dt = DateTime.civil(2000, 1, 1, 0, 0, Rational(1, 1_000_000)) # .000001s
Time.at(dt).usec
# before => 0
# after  => 1

dt = DateTime.civil(2000, 1, 1, 0, 0, Rational(123_457, 1_000_000))
Time.at(dt).usec
# before => 123456
# after  => 123457

Why this is correct

• It brings the DateTime branch in line with the TimeWithZone sibling, which already passes an exact rational timestamp so precision survives the round-trip. The exact fractional second is datetime.to_i + datetime.sec_fraction.
• All other behavior is unchanged, verified across whole-second, offset (incl. DST), microsecond-boundary, and pre-epoch (negative timestamp) values: the resulting instant matches datetime.to_i + datetime.sec_fraction exactly, and the returned Time is still local (getlocal).