Add dft_rast_consensus() for temporal mode smoothing #8

Problem

Single-year land cover comparisons (e.g. 2017 vs 2023) pick up noise from:

• Misclassification in individual years (cloud shadow, spectral confusion)
• Slight geometric shifts between satellite passes (edge pixels flip classes)
• Seasonal timing differences (bare field vs crop depending on overpass date)

This inflates apparent change, especially at class boundaries.

Proposed Solution

Add dft_rast_consensus() — compute per-pixel mode (most frequent class) across multiple years to create a temporal consensus raster.

# Early period consensus: mode of 2017, 2018, 2019
early <- dft_rast_consensus(classified[c("2017", "2018", "2019")])

# Late period consensus: mode of 2021, 2022, 2023
late <- dft_rast_consensus(classified[c("2021", "2022", "2023")])

# Compare consensus periods — much less noise
trans <- dft_rast_transition(list(early = early, late = late),
                              from = "early", to = "late")

Scope

• Input: named list of classified SpatRasters (same extent/resolution) — works with any categorical raster, not LULC-specific
• Output: single SpatRaster with per-pixel mode, retaining factor levels
• Optional: confidence layer (proportion of years agreeing on the mode, e.g. 3/3 = 1.0, 2/3 = 0.67)
• Mean/median don't apply to categorical data — mode is the correct measure

Tradeoffs

• Reduces false positives (misclassification noise) significantly
• Smooths real rapid change (cleared in 2018, regrown by 2019 disappears)
• Requires fetching more years from STAC (e.g. 6-7 instead of 2)
• Confidence layer helps users assess reliability of detected change