Time offset fix - #43

A site the radar never sampled comes back from Py-ART as an all-NaN
column. column_vertical_profile averages an empty set of rays, so
time_offset -- the mean gate time -- is NaN right along with every
field. subset_points purged every all-NaN variable and took
time_offset with it, then raised KeyError: 'time_offset' three lines
later at the gate_time calculation. np.isnan returns True for NaT, so
the check fired on the time columns as readily as on the float ones.

The whole scan was lost, and every co-located site in it, not just the
one out of coverage.

time_offset is bookkeeping, not a measurement. It is now exempt from
the purge and from the dropna height mask, so an out-of-coverage
column comes back whole with time_offset and gate_time as NaT --
missing, which is true.

The dropna exemption also fixes a second, quieter loss: a partly-NaN
time_offset stayed in the dataset and helped decide which heights
survived, discarding gates the radar had genuinely sampled.

Two more in the same block:

- da.drop(v) -> da.drop_vars(v). The former is deprecated and emitted
  a FutureWarning on every all-NaN field.
- Guard the time_offset.drop_duplicates call in the InvalidIndexError
  handler behind the RHI check. time_offset is only bound in the RHI
  branch, so a non-RHI column with duplicate heights hit NameError, or
  silently reused the previous site's offsets.

This predates the pluvio accumulation fix and is independent of it;
the all-NaN purge came in with d837757 (ADD: XSAPR). Scans dropped to
this KeyError were skipped rather than corrupted, so no written output
needs revisiting on its account.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>