fix(arrow/array): support zero-column tables

[fallintoplace]

Rationale for this change

Arrow schemas and record batches may have zero fields while still carrying a logical row count. NewTableFromSlice indexed the first column unconditionally and panicked for an empty schema. NewTableFromRecords inferred rows from columns, so zero-column record batches lost their row counts.

What changes are included in this PR?

Handle the empty column slice without indexing cols[0]. When the schema has no fields, sum RecordBatch.NumRows to preserve the logical row count. Tables with columns keep the existing column-based row inference. Regression tests cover both constructors.

Are these changes tested?

Yes:

• go test ./arrow/array
• go test -race -count=1 ./arrow/array

Are there any user-facing changes?

Zero-column table construction no longer panics, and zero-column record batches retain their logical row counts. Existing behavior for tables with columns is unchanged.

[zeroshade]

Under what scenario would we have zero-fields but a non-zero logical row count? That doesn't make sense to me

[fallintoplace]

The concrete case is projecting or removing all columns while retaining row cardinality. Arrow models the record batch length independently from its fields, so a zero-column batch can still represent N rows. Arrow Go already supports constructing, filtering, serializing, and reading those batches.

[zeroshade]

It's an odd case, but having a clear behavior instead of a panic is still an improvement. Given it's a pretty small change, I'm okay with it.