[BUG] Resampled/CMORised time axis off by ~1 day for pre-1582 dates (GREGORIAN vs proleptic_gregorian)

[rhaegar325]

Summary

The WCRP CMIP6 compliance check [TIME001] Check Time Squareness fails on CMORised
ocean output for ACCESS-ESM1-5, with the time coordinate off by exactly one day:

[TIME001] Check Time Squareness
* Mismatch at index 0: expected 36706.500000, got 36707.500000.
  (table_id=Oyr, frequency=yr, var=no3, midpoint=True)

The numeric time values (and their bounds) are computed in the mixed Julian/Gregorian
"standard" calendar, but the file ultimately declares proleptic_gregorian, so for
dates before the 1582 Julian→Gregorian reform the two disagree by ~1 day.

The same root cause produces a related CF §7.1 finding:

'time_bnds' has attr 'calendar' with value 'GREGORIAN' that does not agree with
its associated variable ('time')'s attr value 'proleptic_gregorian'

Root cause

ACCESS-ESM1-5 model files label their time axis with the non-CF calendar name
"GREGORIAN". Two pieces of code interpret this label inconsistently:

1. base.CMORiser._check_calendar() rewrites "GREGORIAN" → "proleptic_gregorian"
   (a documented TODO workaround for the model's mislabelling). It changes only the
   label, not the numeric values — which is correct, because the model's values are
   genuinely proleptic-Gregorian (the model has no Julian transition).

2. utilities.calculate_time_bounds() and the resampling midpoint helper
   (_shift_resampled_time_to_period_midpoint) read the raw "GREGORIAN" label and pass
   it to cftime, which treats "GREGORIAN" as the CF standard calendar
   (Julian before 1582-10-15). They therefore build the year/period boundaries — and the
   yearly midpoint — in the Julian frame.

For year 0101 this gives:

Quantity	standard / Julian (what we compute)	proleptic_gregorian (what the file declares & the checker expects)
0101-01-01 (days since 0001-01-01)	36525	36524
Yearly midpoint	36707.5	36706.5

The 1-day gap is exactly the Julian-vs-proleptic leap-day difference accumulated over
years 1–101 (Julian counts year 100 as a leap year; proleptic_gregorian does not).

The checker validates against the file's declared calendar (proleptic_gregorian), so it
expects 36706.5 and flags our 36707.5.

Impact

• [TIME001] fails for all CMORised ACCESS-ESM1-5 output whose time axis falls before
  1582 (e.g. piControl/historical spin-up years like 0101).
• CF §7.1 reports a time_bnds vs time calendar mismatch (GREGORIAN vs
  proleptic_gregorian) for the same reason.
• Post-1582 data is unaffected (the two calendars coincide), which is why the
  discrepancy is easy to miss.