There are two blocks: mu and Lvec. The full multivariate approximation would allow for correlations between \mu_i and L_{jk}, and there may be useful relationships to model there (full-rank ADVI is precisly this). For simplicity I broke these parameters into blocks.
The overall rationale is just to deal with the slight chance of there being numerical issues in the decomposition (for instance, if the number of posterior samples is fewer than the number of variables). The specific values are entirely arbitrary; but small with respect to the diagonal of cov(init['mu']) (etc).