sql: allow UDFs in EXECUTE arguments

[mgartner]

We currently disallow UDFs from being used in EXECUTE arguments (see #99008, #108213). UDF calls cannot be simply evaluated like placeholder expressions are currently:

cockroach/pkg/sql/opt/norm/factory.go

Line 355 in f59f0f4

d, err := eval.Expr(f.ctx, f.evalCtx, e.(*memo.PlaceholderExpr).Value)

UDFs must be built by the optimizer and executed as routines. #107767 attempted to address this by replacing placeholders with expressions instead of constants. There were some major issues with this approach though:

1. If the UDF references any database objects, like tables, indexes, columns, or views, then the catalog must be initialized to build the expression. During placeholder assignment, the catalog is not initialized.
2. Inlining the UDF will result in the UDF being effectively evaluated each time the argument is referenced in the prepared statement. For example, if the prepared statement looks like PREPARE p AS SELECT $1::INT + $1::INT, then f will be evaluated twice when running EXECUTE p(f()). This is fine for stable, immutable, and leakproof functions, but causes unexpected behavior for volatile functions. See the related issue sql: EXECUTE arguments should be evaluated only once #108027.

Jira issue: CRDB-30365