fix: string_to_array('', delim) returns empty array for PostgreSQL compatibility

[dd-david-levin]

Problem

string_to_array was returning incorrect results for empty string input — both when the delimiter is non-empty and when the delimiter is itself an empty string. This diverges from PostgreSQL behavior.

Query	DataFusion (before)	PostgreSQL (expected)
string_to_array('', ',')	['']	{}
string_to_array('', '')	['']	{}
string_to_array('', ',', 'x')	['']	{}
string_to_array('', '', 'x')	['']	{}

Results from datafusion-cli

Root cause: Rust's str::split() on an empty string always yields one empty-string element, so "".split(",") produces [""]. Additionally, the empty-delimiter branch unconditionally appended the (empty) string value. This is subtle because Arrow's text display format appears not to quote strings, so [""] renders as [] — indistinguishable from an actual empty array. Using cardinality() reveals the current length is 1, not 0.

PostgreSQL reference: db-fiddle

Fix

In datafusion/functions-nested/src/string.rs:

• Non-empty delimiter (Some(string), Some(delimiter)): added if !string.is_empty() guard to skip splitting when input is empty.
• Empty delimiter (Some(string), Some("")): added if !string.is_empty() guard so the string value is only appended when non-empty.

Both the plain variant and the null_value variant are fixed (4 checks total).

Tests

Added sqllogictest cases in datafusion/sqllogictest/test_files/array.slt using cardinality() to unambiguously verify the arrays are truly empty (not just displaying as empty):

SELECT cardinality(string_to_array('', ','))    -- 0
SELECT cardinality(string_to_array('', ''))     -- 0
SELECT cardinality(string_to_array('', ',', 'x'))  -- 0
SELECT cardinality(string_to_array('', '', 'x'))   -- 0

Each test covers one of the four is_empty guard checks. All four fail without the fix (returning 1 instead of 0).

[LiaCastaneda]

This makes sense to me

[gabotechs]

Looks good. Thanks @dd-david-levin for the PR and @LiaCastaneda for the review! I'll leave this here until tomorrow in case someone else wants to chime in.