[Relax][ONNX] Fix TopK scalar K extraction in from_onnx

[javierdejesusda]

Root Cause

TopK._impl_v11 extracted k with int(k.data.numpy()). ONNX emits K as a single-element 1-D tensor constant, so numpy() returns a 1-D array and int() raises TypeError: only 0-dimensional arrays can be converted to Python scalars, failing conversion of any model with a TopK node.

Solution

Resolve k with get_constant(inputs[1], params) and extract the scalar with .item(), matching the Trilu and Reshape converters in the same file. get_constant also handles k arriving as a parameter when keep_params_in_input=True.

Test Plan

test_topk in tests/python/relax/test_frontend_onnx.py already builds K as a single-element 1-D INT64 constant, so it exercises this path. .item() returns the scalar for both single-element 1-D and 0-d constants.

Issue

Fixes #19571

[gemini-code-assist]

Code Review

This pull request updates the ONNX frontend's TopK implementation to use .item() when converting the k input to an integer, preventing crashes with 1-D tensors. The reviewer suggested using get_constant for consistency and to handle cases where k is a parameter, while also identifying similar patterns in other operators that may require future fixes.

[cchung100m]

LGTM, thanks to @javierdejesusda 😄

[tlopex]

@javierdejesusda Could you resolve the conflict so that we can merge it in?

[javierdejesusda]

Thanks @tlopex. Rebased onto main and resolved the conflict, keeping the reviewed get_constant(inputs[1], params) + .item() extraction.