Dear Fortranners,

conversions between different character kinds using TRANSFER exhibit
inconsistencies that can occur between expr->representation.string
(which is char*) on the one hand, and expr->->value.character.string.

One issue (in target-memory.cc) is easily fixed by simply passing
a conversion flag that was likely forgotten in the past.

The other issue happens in gfc_copy_expr.  Before we unconditionally
converted an existing representation.string to wide char, which is
definitely wrong.  Restricting that code path to default character
kind fixed the problems I could find and produces dumps that looked
fine to me.  Maybe some expert here can find a better fix.

Regtested on x86_64-pc-linux-gnu.  OK for mainline?  Maybe 11-branch?

Thanks,
Harald