Add MinDaysUntilFull, RSquared, Variance, RecentGrowthRatePerDay,
and RecentDaysUntilFull to CapacityForecast struct for Phase 2
of the prediction algorithm.

Expose concrete min/max daily usedFraction deltas so consumers can
show the actual extremes alongside the statistical summary.

A -1 sentinel collides with legitimate negative values — e.g. days
until 80%% full can legitimately be -5 when usage already crossed
the threshold 5 days ago. Use *float64 with nil for "unknown"
so the type system represents unset state cleanly.

Addresses review feedback from klauspost.

Update struct and field comments to reflect the switch from OLS
linear regression to a Kalman filter. No code or type changes.

The Kalman filter is the forecaster; an OLS R-squared offered no
information about the filter's own predictions and only measured the
linearity of the raw data. Drop the field to keep the API focused on
what the filter actually computes.

Per review feedback, embed the unit in the names so readers do not
have to look at the implementation to discover what they hold:

- GrowthRatePerDay  -> GrowthBytesPerDay
- DataPointCount    -> DailySnapshotCount

The doc comments now also clarify that GrowthBytesPerDay is independent
of DailySnapshotCount.

The previous fields expressed the change as a fraction in [-1, +1],
which the reviewer found unclear. Bytes match the other size fields
(CurrentUsedBytes, CurrentTotalBytes, GrowthBytesPerDay) and tell
the user directly how much data was added or freed that day.