[CCCLL]

A compiled SpEL expression that compares double or float operands
with > or >= evaluates to true when one of the operands is NaN.
In interpreted mode, as in Java, the result is false. As a consequence,
the same expression changes its result once it gets compiled, for
example after about 100 evaluations in MIXED mode.

SpelExpressionParser parser = new SpelExpressionParser();
SpelExpression expression = (SpelExpression) parser.parseExpression("1.0d > T(Double).NaN");
expression.getValue();         // false
expression.compileExpression();
expression.getValue();         // true (false with this change)

Operator.generateComparisonCode() always emits DCMPG/FCMPG, which
push 1 if either operand is NaN, so the IFLE/IFLT jump used for
>/>= is never taken. javac uses DCMPL/FCMPL for > and >=,
and DCMPG/FCMPG for < and <=. This change does the same. < and
<= already returned false and keep using DCMPG/FCMPG.

The new parameterized test covers the four operators with NaN on
either side, for double and float. Without the change, the eight >
and >= cases fail once compiled; with the change, all 16 cases pass.
./gradlew :spring-expression:check passes on JDK 25.

The generated code has been like this since the SpEL compiler was
introduced, so 7.0.x is affected as well.