Perf: export Polars String columns directly from Arrow LargeUtf8 buffers

[stewjb]

Summary

• Adds _export_extract_string_obj_arrow() in sbdf_helpers.c: reads raw UTF-8 bytes directly from the Arrow LargeUtf8 values buffer using the int64 offsets buffer — no Python API calls in the hot loop
• Adds _export_vt_polars_string Cython exporter (id=5) that gets raw pointers via PyArray_DATA() on zero-copy numpy views of the Arrow buffers
• Adds set_arrow_string() on _ExportContext to hold the Arrow buffer views alongside the invalids mask
• _export_polars_setup_arrays intercepts Utf8/String/Categorical/Enum dtypes and takes the new fast path instead of falling through to to_numpy()
• Guards that the Arrow type is large_string (int64 offsets); raises SBDFError with a clear message if not

Why this is faster

Previously, export of a Polars string column did:

1. series.to_numpy() → one Python str object allocated per row
2. C helper: PyObject_Str() + str.encode("utf-8") per element — redundant since Arrow already stores UTF-8

Now it does:

1. series.to_arrow() → zero-copy Arrow array (Polars shares memory)
2. np.frombuffer(buf) → zero-copy numpy views of offsets and bytes buffers
3. C function: sbdf_str_create_len(values_buf + offsets[i], len) — direct slice, no Python objects

Benchmark (psutil RSS, 7 reps, warmup discarded)

100,000 rows, String no nulls

	time (ms)
Export: pandas df	58.3
Export: polars df (old: via to_pandas())	71.4
Export: polars df (new: Arrow direct)	25.7

100,000 rows, String ~10% nulls

	time (ms)
Export: pandas df	88.4
Export: polars df (old: via to_pandas())	105.2
Export: polars df (new: Arrow direct)	36.6

The Arrow direct path is ~56% faster than the pandas baseline and ~64% faster than the old polars workaround. The remaining time is dominated by sbdf_str_create_len (one malloc + memcpy per string — unavoidable in the current SBDF format).

Edge cases

• Categorical/Enum: cast to Utf8 before to_arrow(), same as existing path
• All-null or empty series: bufs[2] may be None; falls back to np.empty(0, uint8) — C loop skips all rows via invalids mask
• ChunkedArray (older Polars): combine_chunks() called if to_arrow() returns a chunked result

Test plan

• All 53 existing SBDF tests pass (pytest spotfire/test/test_sbdf.py)
• Run benchmark and confirm string export time decreases vs prior baseline
• Verify Categorical/Enum string columns export correctly (covered by existing polars write tests)
• Verify null string columns export correctly (covered by test_write_polars_nulls)

🤖 Generated with Claude Code