[BUG] regexp_replace with back references produces incorrect results on GPU

[andygrove]

Describe the bug

Spark supports back references in regexp_replace but this functionality does not work correctly on the GPU.

Example

val df = List("a1","b22","333").toDF("str").repartition(2)
val df2 = df.withColumn("repBackRef",regexp_replace('str,"(\\d+)$",";$1"))
df2.show

Expected Results

+---+----------+
|str|repBackRef|
+---+----------+
| a1|       a;1|
|b22|      b;22|
|333|      ;333|
+---+----------+

Actual Results

+---+----------+
|str|repBackRef|
+---+----------+
| a1|      a;$1|
|b22|      b;$1|
|333|       ;$1|
+---+----------+

Steps/Code to reproduce bug
See above.

Expected behavior
We should fall back to CPU, or produce the correct results.

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

Additional context
None