UDF fails when subexpression elimination is used in interpreted mode

[chris-twiner]

As UDF expects compilation always this fails when a child cannot compile, such as ExpressionProxy.

If, in interpreted mode, a child is deemed an identical subexpression (as occurs in the UdfTests."multiple two argument udf" test) spark is able to swap the contents as an optimisation to reduce re-calculation.

In the case where the analyser deems compilation the correct approach this isn't an issue. Targeted fix as part of #800.

Known to occur on Databricks 11.3 and other lower runtimes, fixed as part of 3.4 in SPARK-41991 .

[chris-twiner]

changing title to represent the wider issue (not fixed in b16067). The above fixes subexpression elimination used in interpreted mode.

This problem exists for all such code, and any could be added. In order to properly fix it needs an eval mode as well.