[DiadkoShmek]

Fixes #20882.

PyTorch exports the grid in [N, H_out, W_out, 2] order. With layout="NCHW", Relax reads it as [N, 2, H_out, W_out]. Passing it through directly makes the output shape (1, 1, 5, 2) for a grid whose expected output is (1, 1, 3, 5). A 2×2 grid happens to yield the expected shape but can still sample the wrong coordinates.

Move the coordinate axis before emitting the Relax operator. Use a non-square grid in the IR test and compare compiled LLVM results against PyTorch eager for both grid shapes.

Local validation on 23a6dbe, Linux x86-64, PyTorch 2.10.0, and LLVM from TIRx 0.1.0.

• On the base, the IR test and both numerical cases fail. The square case differs in three of four values.
• With the patch, the grid_sample and adjacent affine_grid subset passes 5 tests. Both end-to-end outputs match PyTorch eager.
• Ruff lint and formatting checks pass for the changed files.

Codex assisted with the patch and regression tests.