onnxruntime: Fix cuda-toolkit version range

[danimtb]

Summary

• Version range in onnxruntime CUDA requirement from cuda-toolkit/[>=12 <14] to cuda-toolkit/[>=12 <13.3] so the sample no longer resolves to cuda-toolkit/13.3.1.
• Workaround for an nvcc 13.3 host-parser regression that breaks the CUDA EP build when compiling moe_kernel.cu against Abseil container headers.
• Unrelated to other sample/recipe branches; regressions started after cuda-toolkit/13.3.1 landed (Add cuda-toolkit/13.3.1 #8, ~2026-08-04). Earlier green runs used cuda-toolkit/13.2.x with the same abseil/20260107.1.

Root cause

With with_cuda=True, Conan resolves cuda-toolkit/[>=12 <14] to 13.3.1. Building ONNX Runtime’s MoE CUDA kernel fails in nvcc’s EDG/cudafe++ host path with errors like:

error: a using-declaration may not have the 'typename' keyword
... using template type parameter 'absl::...::IfRRef<...>::AddPtr<K>' after 'typename'

See failure at https://github.com/conan-io/cuda-conan/actions/runs/31498690863/job/94037761221

Seen in Abseil headers such as:

• absl/container/internal/raw_hash_map.h
• absl/container/internal/btree_container.h

Same Abseil + recipe combination builds with CUDA 13.2.x, so this is treated as an nvcc 13.3 incompatibility rather than an Abseil/recipe bug.

Fix

In samples/onnxruntime/all/conanfile.py:

• self.requires("cuda-toolkit/[>=12 <13.3]")
• Short comment documenting the nvcc 13.3 / Abseil macro issue

tool_requires("cuda-toolkit/<host_version>") stays unchanged and follows the host requirement (resolves to 13.2.2 today).

Test plan

• Samples CI: onnxruntime with CUDA builds successfully (resolves cuda-toolkit/13.2.2)
• Confirm graph does not select cuda-toolkit/13.3.1

See https://github.com/conan-io/cuda-conan/actions/runs/31582068953

TODO

• Revisit this when NVIDIA/Abseil fix the nvcc 13.3 parse issue (or pin a known-good Abseil/CUDA pair)