Hello,

this workarounds the regression I introduced with the fix for pr104228.
The problem comes from the evaluation of character length for the 
associate target (y) in the testcase.
The expression is non-scalar which is not supported at that point, as no 
scalarization setup is made there.

My initial direction to fix this was trying to modify the evaluation of 
expressions so that the rank was ignored in the context of character 
length evaluation.
That’s the patch I attached to the PR.
The patch I’m proposing here tries instead to avoid the need to evaluate 
character length through sharing of gfc_charlen between symbols and 
expressions.
I slightly prefer the latter direction, though I’m not sure it’s more 
robust.

At least it passes regression testing, which proves some basic level of 
goodness.
OK for master?
Even if PR104570 is a 12 regression only, the PR104228 fix is targeted 
at all the open branches so this one as well.
OK for 11/10/9?