I think best way to generated Heteroscedastic data would be, first
fix or assume some DGP of your choice, you may assume any arbitrary
model parameter(s) provided forms of those parameters are legitimate
under the model space. Then just simulate numbers starting from the
Residual (the form of that residual you already must have assumed).
Given some initial value(s), you would then get some Time series. You
may chose to generate a lengthy TS as most of the parameters
estimations happen within Asymptotic concept. Then apply your
statistical test on that realized series.
You may repeatedly generate TS and check the distribution of estimated
model parameters etc.