Fix col_reduce_longcolumn for negative-stride views

[mateuuszzzzz]

Fixes: #4528

Details

The thread coordinate (lid.x) in col_reduce_longcolumn is an unsigned integer and causes implicit conversion of in_idx to an unsigned integer in the expression in_idx + lid.x. This implicit conversion wraps the input offset for negative-stride views. PR fixes this by adding an explicit type cast of lid.x to IdxT, ensuring both operands share the same signed type and avoiding the implicit conversion of in_idx.

Tests cover operations relying on reduce kernels and all dispatch paths for the col_reduce* family.
I verified that all test cases picking col_reduce_longcolumn fail on main and pass after applying the fix.

• ☑️ I understand it is strictly prohibited to use AI to write PR description

• AI usage disclosure: Opus 5 assisted with:

  • identifying what other functions except sum are impacted by the bug (and pointed me to col_reduce* family)
  • writing tests that cover dispatches to all col_reduce* kernels
  • checking that current fix does not introduce any regressions.

  I verified the exact reproduction conditions, implemented a fix, and ran local checks and tests before pushing.