[Fix][DLight] Localize private scalar reduction buffers

[SamJSui]

DLight's scalar-output path returns before the normal reduction-buffer
scope assignment. For length-one argmin, the paired index and value
temporaries therefore remain global. The index is read by the final
result block and must remain global, while the value temporary ends in
the producer kernel and becomes an illegal internal global allocation
during CUDA code generation.

This change assigns local scope only to root-allocated global buffers
that are produced under unit-extent loops and are not accessed by
another block. The index temporary therefore retains global scope.

A focused regression covers the length-one case and verifies that
non-unit reductions retain global scope.

Fixes #20060.

Testing:

• python -m pytest tests/python/s_tir/dlight/test_gpu_general_reduction.py -q — 8 passed
• python -m pytest tests/python/s_tir/dlight -q — 101 passed, 1 skipped, 1 xfailed
• Adjacent Relax tests — 52 passed
• Exact issue reproducer — passed
• Changed-file pre-commit hooks — passed