Fix CLI register path dedup and local timm model export (#2417, #2423) - #2478

What does this PR do?

Fixes two issues addressable in the core optimum repository and updates the contributing guide to reflect the current split into subpackages.

#2417 — sub-commands fail on RHEL (lib64 → lib symlink)

load_optimum_namespace_cli_commands deduplicated submodule_search_locations with a plain set() over the raw strings. On systems where lib64 is a symlink to lib (e.g. RHEL), this returns two entries pointing to the same physical directory, so subpackage commands were registered twice and spurious __init__.py warnings were emitted.

Fix: deduplicate by the resolved physical path (Path.resolve()) and guard iterdir() with a try/except OSError.

#2423 — cannot export local timm models

The timm loading branch always passed hf_hub:{model_name_or_path} to timm.create_model. For a local directory, timm interprets the hf-hub/hf_hub source prefix as a Hub repo id and fails to download it. timm does not accept a bare local path — the local-dir: source prefix is required for local folders.

Fix: select the source prefix based on whether the argument is a local directory:

• local directory → local-dir:{path}
• otherwise → hf_hub:{path} (unchanged)

CONTRIBUTING.md

Added a "Where does my contribution belong?" section mapping features/backends to their subpackage repositories (optimum-onnx, optimum-intel, optimum-habana, optimum-neuron, optimum-amd, optimum-nvidia, optimum-quanto, optimum-furiosa, optimum-executorch), and clarifying that this repository holds the shared core (CLI, optimum/exporters/tasks.py, utils).

Tests

• tests/cli/test_cli.py::TestCLI::test_load_namespace_cli_commands_dedup_symlink — regression test for Cannot run sub-commands on RHL because optimum mishandles the lib64 symlink to lib #2417 using an isolated temp directory plus a symlink duplicate and a namespace ModuleSpec.
• tests/exporters/common/test_timm_loading.py — verifies the local-dir: prefix is used for a local path and hf_hub: is preserved for a Hub repo id (mocks timm.create_model, no network/weights).

Both pass with transformers 4.57.6 and transformers 5.15.0.

Notes from triage (no code change needed in this repo)

• Patch release to support transformers >= 5.0.0 #2432 (transformers >= 5.0): the core is already compatible — all AutoModelFor* loaders referenced in tasks.py exist on transformers 5.15, the version gate at tasks.py:156 picks AutoModelForImageTextToText, and setup.py has no upper cap on transformers. Remaining blockers are in the subpackages (e.g. optimum-onnx pins transformers < 5).
• export Qwen3-ASR-0.6B to onnx error #2474 (Qwen3-ASR ONNX export): root cause is a missing qwen3_asr ONNX config, which lives in optimum-onnx.
• RuntimeError converting Whisper model #2338 / I follow the Documentation and I still can't fully export an ASR model to ONNX :( #2434 / Exporting Mistral to ONNX Model #2312 / not able to convert DeepSeek-R1 into Onnx using optimum-cli #2215: not reproducible on current versions with optimum[onnx] + onnxruntime (openai/whisper-tiny and sshleifer/tiny-gpt2 export cleanly); these belong to the optimum-onnx export layer.