[NishchayMahor]

Fixes #2677.

In that issue the reporter takes a 48-month series, shifts every date forward by one
year, leaves y untouched, and the 12-month forecast total moves by -25%. One
point flips sign. @Fediuld found that yearly_seasonality=False makes it go away, and
@gboulmier then asked the question this PR answers: why would turning off a seasonality
the data clearly has make the fit more stable?

Why it happens

Monthly data has one sample per month, so a yearly cycle is sampled 12 times. Above the
Nyquist limit of that spacing, floor(365.25 / (2 * 30.4)) = 6 harmonics, the higher
harmonics alias onto lower ones. Their coefficients are then set by the prior rather
than by the data, so the fitted shape can move a long way for a small change in the
input. yearly_seasonality='auto' asks for order 10 regardless of spacing, so four of
those ten harmonics are unidentifiable.

Prophet already guards the other two seasonalities on exactly this basis. forecaster.py
declines weekly once min_dt >= Timedelta(weeks=1) and daily once
min_dt >= Timedelta(days=1). Yearly has no min_dt term at all. That inconsistency is
the argument here, not a general claim about what users do.

The change

set_auto_seasonalities caps the default yearly order at the Nyquist limit of the
observed minimum spacing. Weekly and daily drop the component entirely at that point;
yearly keeps the component and lowers the order instead, so monthly data still gets a
yearly shape. An explicitly requested order is left alone.

spacing	order before	after
hourly	10	10
daily	10	10
weekly	10	10
monthly	10	6
quarterly	10	2
annual	10	0

Effect on the reported series

Reporter's own data, dates shifted one year, y unchanged, uncertainty_samples=0:

	before	after
12-month total	-25.1%	-0.3%
5th point	-104.5% (sign flip)	+10.3%
7th point	+56.4%	+4.6%
min multiplicative term	-1.0535	-0.4688
negative forecast produced	yes	no

The multiplicative term below -1 is what produced the sign flip: yhat = trend * (1 + m).

Test suite: 139 passed, 6 skipped, against a 138/6 baseline I measured on 5cf15c0.
The new test asserts the order for six frequencies plus the two override paths, and
fails on current main.

What I did not verify, and the tradeoffs

• min_dt is the minimum gap, not the typical one. A monthly series with one stray
  1-day gap keeps order 10 and is unhelped. I verified that. It is the fail-safe
  direction and it matches how the weekly and daily guards already behave, but median
  spacing would bind more often if you would rather have that.
• Monthly users' numbers change. Anyone on auto with monthly data gets a different
  yearly shape after this. An logger.info says so and names the override. If you would
  rather not move existing fits, the warning alone (no cap) is a one-line version of this
  PR and I am happy to cut it back to that.
• This does not fix Unexpected change in predictions on close to identical data #2676. Different cause.
• I have not benchmarked accuracy across a corpus of real monthly series. The claim here
  is identifiability and internal consistency, not that forecasts get better on average.
• Two docs lines describing yearly order 10 as the default become qualified for coarse
  data. I left the docs alone pending a call on the approach.