Fix layout inference for column-vector tensor views

[TaoTao-real]

Fixes #190.

What was wrong

For some 2D tensor views (e.g. shape (16 x 1) with contiguous strides (1, 1)), the emitted GlobalTensor<...> layout parameter could be inferred as ND even though the expected layout is DN (column-vector).

What this PR changes

• Infer layout using a 5D right-aligned shape/stride view (pads leading dims with 1) to match the GlobalTensor ABI.
• Add a vector-friendly rule: when minor dims are ambiguous (e.g. one dim is 1), prefer:
  • DN for (rows x 1)
  • ND for (1 x cols)
• EmitC lowering now prefers the layout attribute inferred/propagated in IR, instead of re-inferring layout in multiple places.
• Enable the layout inference pass by default (can still be disabled via --disable-infer-layout).

Tests

• Added a regression sample test/samples/Layout/tensor_view_infer_layout_dn.py + guard in test/samples/runop.sh.