fix(datafusion): report an unconvertible column statistic as absent

[jackylee-ch]

stats_set_to_df (vortex-datafusion/src/convert/stats.rs:30-66) ran each statistic through two
vortex_expects. Both sides of that conversion come from the file — the stored value out of the
footer's stats set, the dtype off the column — so both can legitimately fail on a file we did not
write:

• Stat::dtype returns None where the statistic does not apply:
  vortex-array/src/expr/stats/mod.rs:173,175 return None for Max/Min of a DType::Null
  column, and Sum::return_dtype does the same for string, binary, list and struct columns. That hit
  vortex_expect("must have a valid dtype").
• Scalar::try_new rejects a stored value that does not match the column dtype, hitting
  vortex_expect("Stat::Minsomehow had an incompatibleDType").

Either one panicked inside DataFusion planning instead of returning an error or simply forgoing the
statistic.

Fix

Forgoing it is what the end of the same chain already chose: try_to_df().ok() drops a scalar
DataFusion cannot represent. This extends that to the whole conversion, which also lets the three
near-identical blocks collapse into one helper — and is what the TODO(connor) above them
("There's a lot that can go wrong here, should probably handle this more gracefully") asked for, so
that comment goes too.

The statistic's Precision is preserved rather than flattened: an exact statistic stays
Precision::Exact, and only the unconvertible ones become Absent. That distinction is easy to lose
here, since StatsSet::get returns Vortex's Precision and not an Option.

Tests

cargo test --release -p vortex-datafusion: 341 passed, 335 before. Four cases cover the
unconvertible shapes — Min/Max of a null column, Sum of a string column, and a stored value
that disagrees with the column dtype — and two more pin that a convertible statistic keeps its exact
or inexact precision.

Restoring the vortex_expects fails all four unconvertible cases, on both panic sites
(vortex-error/src/lib.rs:666 for the Option and :659 for the Result), and leaves the
precision and distinct_count cases green.

AI assistance

Written with agentic AI assistance; my first version returned Option instead of Precision and
would have silently flattened every statistic's precision — the compiler caught it, and the two
precision cases above exist so a reviewer does not have to.