v3.5.0

Radiation morphs change: diffuse horizontal and direct normal now keep the
EPW's own split, and EPWs without extraterrestrial radiation morph correctly.
Global horizontal radiation and every other variable are unchanged.

Fixed

• No direct normal from EPWs with empty extraterrestrial columns. The
  radiation morph read exthorrad_Whm2 for the clearness index and capped
  direct normal at extdirrad_Whm2. weather-file-builder up to 2.3.0 wrote
  both as 0 in ERA5 typical and extreme years, so every Radiation or Clouds
  and Radiation morph of such a file came out diffuse only, with direct
  normal 0 in every hour. When a column holds no positive value it is now
  computed from the solar position (new procedures.extraterrestrial_or_computed,
  solar.extraterrestrial_radiation) and a warning is logged. The computed
  columns are also written to the morphed EPW (rounded, 0 at night) and named
  in COMMENTS 2. Filled columns are used and written as they are. calc_dirnor also falls back to the solar constant
  when its ceiling column is empty.

Changed

• Diffuse and direct normal keep the EPW's split (procedures.morph_radiation_split).
  Before, both were recalculated from scratch with the Ridley et al. (2010)
  model, which differs from the model the EPW was split with (weather-file-builder
  uses DIRINT). That moved diffuse and direct normal by about 5% even with no
  climate signal (an ERA5 TMY at Zurich: diffuse -5%, direct normal +5%), and
  the error went into every future file. Now the diffuse fraction is the EPW's
  own plus the change the Ridley model predicts between present and morphed
  global radiation, bounded so it never passes the model's future value, and
  direct normal scales the EPW's own value by the change in horizontal beam
  (beam change projected mid-hour where the present beam is under 10 W/m2).
  With no climate signal both are now returned unchanged, and the identity
  tests require it. Zurich ERA5 TMY, CH2025 GWL 3.0 median (global +2.0%):
  diffuse -1.3%, direct normal +4.3% (3.4.0 with the columns filled: -7.7%
  and +10.4%).
• Radiation solar geometry is taken mid-hour. EPW rows are hour-ending
  and the dataframe index marks the start of each hour; the radiation
  procedures took the sun's position at that start, 30 minutes early. They
  now use the middle (solar_geometry(..., offset_minutes=HOUR_MIDPOINT_MINUTES)).
  The returned index and the EPW's date and time columns are unchanged.
• calc_difhor and calc_dirnor remain for a full re-split but the morph no
  longer calls them.