[BUG] regexp_replace [^a] has different behavior between CPU and GPU for multiline strings

[andygrove]

Describe the bug

With the pattern [^a] (meaning match any character except for 'a'), cuDF does not match newline characters but Java/Spark does, resulting in different behavior as demonstrated by these code snippets.

Spark/Java behavior

scala> import java.util.regex._
scala> val p = Pattern.compile("[^a]")
scala> p.matcher("a\nb\nc").replaceAll("1")
res0: String = a1111

cuDF behavior

>>> import cudf
>>> cudf.Series(['a\nb\nc']).str.replace('[^a]', '1', regex=True)
0    a\n1\n1

Steps/Code to reproduce bug

Add the following integration test to string_test.py

def test_regexp_replace_multiline():
    gen = mk_str_gen('[abcd]{0,3}\n[abcd]{0,3}')
    assert_gpu_and_cpu_are_equal_collect(
            lambda spark: unary_op_df(spark, gen).selectExpr(
                'regexp_replace(a, "([^a])|([^b])", "1")'),
            conf={'spark.rapids.sql.expression.RegExpReplace': 'true'})

Expected behavior
We either need to document the difference in the compatibility guide or we need to find a way to make the behavior consistent.

Environment details (please complete the following information)
N/A

Additional context
None