The first thing to check is that the data consumer (which the OP notes is not them) may be able to actually process the upper-triangular matrix as-is; look at LAPACK, which has routines that operate on triangular matrices directly without needing to copy - in other words, this task may not actually be needed.

If it is, the approach I would take using PDL is to pre-allocate a correctly-sized disk-based ndarray (so that RAM is not a limitation) using PDL::IO::FastRaw (untested), from perldl:

use PDL::IO::FastRaw;
$pdl = mapfraw('fname', {Creat => 1, Dims => [460_000, 460_000], Datat
+ype => double()});
[download]

It looks like currently PDL::IO::Misc (with rcols) and PDL::IO::CSV don't have a facility to read into an existing (possibly disk-based) ndarray. It would still be possible, albeit slow, to read the file line-by-line, probably using Text::CSV_XS, using slice to insert each row into the ndarray.

Once the ndarray exists, to copy one triangle into the other would be super-easy using PDL::LinearAlgebra::Real#tricpy.