Get text positions from transformed text

Using the non-transformed text is bad because the
transformed text is what's actually on the screen,
so it is the thing whose position is relevant.

Bug: 69044691
Test: Updating a11y cts test to use an all-caps
german strong s, which causes the transformed text
to be longer than the original text. That changed
test fails without this fix.

Change-Id: Ia7aac3b4322ffb48f10771f4d1b33c80ca28b120