Kruskal-Wallis is quite robust and loss of power is not that big even if the distribution is symmetrical. Remember that if you want to be even more robust you can always use bootstrap.

I am not sure I follow you... Power has largely to do with sampling and variability, whereas regression to the mean concerns the fact that, on average, cases tend to be close to the average, and not to the extremes.

RevMan provides you RR, not logRR. You thus need to log-transform it (natural logarith). If the SE is not reported, you can back calculate it from the 95% confidence interval, which is indeed built as: exp ( logRR +/- 1.96 * SE ).

I have edited my reply providing some suggestions for beginners. Note though that if you want to submit it for a scholarly publication you will need to do it proficiently.