add KL divergence backward helion kernel [attempt 2]

[williamwen42]

#802 but without ghstack

[oulgen]

• fix the lint error
• add the backwards implementation to benchmarks/run.py -- there's a big list, grep for rms_norm
• add a unit test to test_examples.py

[williamwen42]

@oulgen I had to type a number of my kernel arguments with hl.constexpr instead of the actual types in order to prevent compilation errors, e.g. typing compute_y_true_grad as bool would result in a branch type divergence error; typing eps as float would result in a GuardOnDataDependentSymNode when attempting to compute math.log(eps). Can you advise if there's a more preferred way for me to express the computation?