Fix EarthSatellite.epoch float-day truncation

[shaneraphel]

Fixes #1141.

EarthSatellite.epoch is documented as the TLE epoch. The current path packs satrec.epochdays (and the from_satrec Julian-date fraction) into a single float day argument of ts.utc, which truncates the fraction. That offset is up to ~34 µs and moves a LEO evaluated at its own epoch by ~25 cm. from_omm already uses ts._utc_jd and is not rewritten.

Pass the day fraction as seconds:

        whole_days = int(satrec.epochdays)
        seconds = (satrec.epochdays - whole_days) * DAY_S
        self.epoch = ts.utc(year, 1, whole_days, 0, 0, seconds)

and the same on from_satrec. Leap seconds are not rewritten.

Measured on skyfield 1.55 / sgp4 2.27, TLE 55910 23076.01808579, builtin timescale:

constructor	float-day offset	seconds form
EarthSatellite(line1, line2)	−12.35 µs (93.4 mm TEME)	~0 µs (~0 mm)
from_satrec	−12.35 µs	~0 µs
ts._utc_jd (OMM control)	0	0

[brandon-rhodes]

Thanks for looking into this, I'll hopefully get the chance to look at this in detail within the next week!

[brandon-rhodes]

Since d618ee9 fixed the underlying issue with a slightly different approach, I'm going to close this PR. But thanks for the work to find a fix!