potentially improve warning/error messages

[clarkliming]

related to #387 but still a little different

in some scenarios the variance estimate is 0 if the values match so well to the fixed effect, leading to a incorrect hessian/gradient

see the following example

fit <- mmrm(FEV1 ~ RACE + SEX + ar1(AVISIT | USUBJID), data = fev_data)

fev_data$FEV1[!is.na(fev_data$FEV1)] <- fitted(fit)

mmrm(FEV1 ~ RACE + SEX + ar1(AVISIT | USUBJID), data = fev_data)

we might improve this error messages to be clearer

[danielinteractive]

@clarkliming What would be the approach here? E.g. can you define the condition for a better error message based on a 0 or almost 0 variance estimate?

[clarkliming]

probably in defining the objective, checks the values of sd and if they are smaller than a specific value (like log(1e-16)) we then raise an error with suitable message