fix(fast-path): preserve Arrow type when building new-column table

[FANNG1]

Fixes #40

Problem

FastPathFragmentWriter built the new-column Arrow table via:

arr = pa.array(s.to_pylist() if hasattr(s, "to_pylist") else list(s))

pa.array(s.to_pylist()) loses Arrow type information. Python floats are always float64 and Python lists carry no fixed-size constraint, so fixed_size_list<item: float>[N] (e.g. an embedding vector from a @daft.func.batch(return_dtype=DataType.fixed_size_list(DataType.float32(), 128)) UDF) was silently widened to list<item: double>. Lance then commits the wrong type in the schema.

Fix

Use s.to_arrow().combine_chunks() instead, which preserves the Arrow type that daft declared in the UDF's return_dtype.

if hasattr(s, "to_arrow"):
    arr = s.to_arrow()
    if isinstance(arr, pa.ChunkedArray):
        arr = arr.combine_chunks()
else:
    arr = pa.array(list(s))

Test

TestRegressions::test_fixed_size_list_float32_type_preserved — creates a UDF returning fixed_size_list<float32>[8], merges via fast path, and asserts the Lance schema preserves the type and values are non-null.