fix(export): resolve timm image_size from pretrained_cfg

[vortex-captain]

Summary

• Fix winml config emitting [1, 3, 64, 64] instead of [1, 3, 224, 224] for timm image-classification models (e.g. timm/repghostnet_200.in1k).
• Root cause: TimmWrapperConfig stores shape info in pretrained_cfg["input_size"] (a plain dict). Optimum's NormalizedConfig only walks PretrainedConfig children, so the value is invisible to DummyVisionInputGenerator, which then defaults to 64x64. timm models also have no preprocessor_config.json on the hub, so winml's existing fallback misses too.
• Fix: when preprocessor_config.json is unavailable, synthesize a preprocessor-style dict from hf_config.pretrained_cfg.input_size. The existing size-parsing block (size int / dict-with-height-width / shortest_edge) is unchanged — the timm concern is isolated at the data-fetch boundary in _get_preprocessor_dict / _synthesize_preprocessor_dict.
• Registry: added timm/mobilenetv3_small_100.lamb_in1k and timm/repghostnet_200.in1k to models_all.json; both PASS perf on CPU. Perf output confirms pixel_values [1, 3, 224, 224] float32.

🤖 Generated with Claude Code

[timenick]

CI failure here is unrelated to this PR — it's network-dependent unit tests

The failing job (run 26930137132) is erroring on tests/unit/export/test_io.py, not on the timm change in this PR. All 9 errors are the same OSError:

OSError: We couldn't connect to 'https://huggingface.co' to load the files,
and couldn't find them in the cached files.

Root cause: two module-scoped fixtures pull configs from the HF Hub at collection time:

• t5_config → AutoConfig.from_pretrained("google-t5/t5-small") (test_io.py#L700-L705)
• qwen_config → AutoConfig.from_pretrained("Qwen/Qwen3-0.6B") (test_io.py#L708-L713)

These fixtures and the T5/Qwen test classes were introduced in #334 (feat: multi-model support with KV cache (T5, Qwen, Mu2), commit 38775de8).

A unit test shouldn't require network access — it makes the suite fail on offline runners, behind firewalls, or whenever the Hub is unavailable, which is exactly what hit this PR.

Suggested fix: build these configs synthetically, the way the Marian/BART tests in the same file already do (e.g. marian_config constructs MarianConfig(...) directly with small dims, no network). T5/Qwen can do the same. If a real from_pretrained download is genuinely needed, those cases belong under tests/integration/, not tests/unit/.

Generated with Claude Code