feat(datafusion): report measured widths for variable-width scan columns

[vivek-bharathan]

A schema fixes no per-row width for strings, binary, lists, maps or
dictionaries, so scans report no byte size at all for any row containing one
and DataFusion falls back to row counts.

Measure what those columns actually cost on the batches a scan has already
decoded, using get_array_memory_size -- the same quantity DataFusion reads
from the other side of a join. A completed scan records a width per column
into the session cache, keyed by dataset version, and plan construction folds
what earlier scans measured into the Lance read nodes.

[lance-gatekeeper]

✅ Gate recommendation: approve.

The completed-scan measurement now reaches variable-width structs and fixed-size lists, while wholly fixed-width layouts retain schema-based sizing. The previously reported join-size underestimates are addressed, and the focused width regressions pass.