Can you please provide the original PROCESS output? How was gender coded? I would suspect, without seeing the data, that you coded gender 1 and 2, but the conditional effect of body image (there are no main effects in moderation anymore, all are conditional!) is when gender has the value 0, which is not a meaningful value then. Therefore, the effect is exactly the amount of .19 smaller as the interaction effect (2.66- .19 = 2.47). So do not refer to the zero order predictors as main effects, that is simply not correct.

Yes, this is possible. Look at your standard errors. The SE for the whole sample is 2.66 but for the separate subsamples they are 1.27 and .80. There is a gender main effect meaning for one gender you have higher scores on the DV than for the other gender. When you test the relationship of body image on self-esteem, the SE is higher for the whole sample because it includes both variance within each gender group, and variance (between-group) due to gender group differences. This is not a moderator effect, but is an effect that happens when subjects are nested in groups that differ. When you compute the separate regressions, you are controlling for group mean differences, which in this case results in smaller standard errors. When group differences are controlled, results can be different from the uncontrolled case, but not necessarily in the way you find here.

Looking are the individual regressions, it doesn't look like there is much of a moderator effect as the coefficients (slopes) are not very different between the two sub-groups, but you would have to look at the statistical significance test to be sure.

Thank your very much for your kind help! :)