Fix percentile/probability message key collisions for PDT 9 and 10

[jbeezley]

This contains a few edge cases related to obscure GRIB2 templates including probabilities and percentiles from ensemble forecasts.

[jbeezley]

I had to make some editorial decisions on how to structure the xarray dataset. Here are the conventions I chose, let me know if you would like to modify anything.

Percentile (PDT 10)

Messages with the same variable but different percentile values are joined along a percentile coordinate dimension.

ds.tmp                    # (time, percentile, latitude, longitude)
ds.percentile             # [1, 50, 99]

ds.tmp.sel(percentile=50) # (time, latitude, longitude)

Probability — split by type (PDT 9)

Messages with different ProbabilityType values become separate variables with the type abbreviation appended.

ds.tmp_prob_abnorm        # P(above normal)   → (time, latitude, longitude)
ds.tmp_prob_bnorm         # P(below normal)   → (time, latitude, longitude)
ds.tmp_prob_nnorm         # P(near normal)    → (time, latitude, longitude)
ds.tmp_prob_below         # P(X < limit)      → (time, latitude, longitude)
ds.tmp_prob_above         # P(X > limit)      → (time, latitude, longitude)
ds.tmp_prob_between       # P(lo < X < hi)    → (time, latitude, longitude)
ds.tmp_prob_between_inc   # P(lo ≤ X ≤ hi)    → (time, latitude, longitude)

# Each carries a probability_type attribute:
ds.tmp_prob_abnorm.attrs['probability_type']
# → "Probability of event in above normal category"

Probability — threshold dimension (PDT 9, same type, varying limits)

When multiple messages share the same variable and probability type but differ by threshold limits, they are joined along a threshold coordinate.

ds.tmp_prob_between_inc   # (time, threshold, latitude, longitude)
ds.threshold              # [1.0, 3.0, 10.0]

ds.tmp_prob_between_inc.sel(threshold=3.0)  # (time, latitude, longitude)

[jbeezley]

Another bug fix for probability thresholds. It's a little ugly, but avoids tuples in coordinates, which always causes issues.

 # Each (lower, upper) pair is a separate variable
 ds.tmp_prob_between_inc_1_3   # (time, lat, lon) — lower tercile
 ds.tmp_prob_between_inc_1_5   # (time, lat, lon) — lower quintile
 ds.tmp_prob_between_inc_1_10  # (time, lat, lon) — lower decile
 ds.tmp_prob_between_inc_3_3   # (time, lat, lon)
 ds.tmp_prob_between_inc_5_5   # (time, lat, lon)
 ds.tmp_prob_between_inc_10_10 # (time, lat, lon)

Single-limit types still use the threshold dimension:

# P(X < threshold) with varying thresholds → threshold dim
ds.tmp_prob_below  # (time, threshold, lat, lon) with threshold=[273, 283, 293]

[mpiannucci]

Just some conflicts since i merged the others first

[jbeezley]

Thanks! Rebased