You can use grouptransform with an anonymous function that calls filloutliers. Let's use your sample data.
group = repelem(['A' 'B'], 1000)';
val = repelem(1:1000, 2)';
t = table(group, val);
This grouptransform call uses the variable group from the table t as the grouping variable. The anonymous function is the same as what you used and Walter each used in your for loops, though I chose to replace it with the double NaN rather than the text 'NaN' like Walter did.
t2 = grouptransform(t, 'group', ...
@(x) filloutliers(x, NaN, 'percentiles', [0 99]));
Let's see what values of val in t were replaced by NaN in t2.
By the way you built t, those do look like the top 1% of values for each group.