fix(openai): infer model input modalities to prevent image read failures

[amitksingh1490]

Summary

Fix OpenAI model modality discovery by inlining a curated model catalog (including input_modalities) so vision-capable models can be selected reliably and image reading no longer fails.
This avoids depending on the OpenAI Models API for modality metadata, which is currently missing/incomplete.

Context

The OpenAI Models API does not consistently return modality information, which breaks Forge’s ability to detect image-capable models and causes image read flows to fail.

Changes

• Replaced the OpenAI provider models endpoint URL with an inline model catalog that includes modality metadata.
• Added input_modalities entries (e.g. text, image) to ensure the UI/selection logic can reliably choose vision-capable models.

Key Implementation Details

• The OpenAI provider entry in provider.json now contains an explicit models array rather than a remote URL.
• Each model includes context_length, tool-call capabilities, and input_modalities to drive capability-aware selection.

Use Cases

• Selecting a model for image reading now works because vision-capable models are explicitly marked with input_modalities: ["text", "image"].
• Local/offline workflows no longer require an additional API call to enumerate models.

Testing

cargo test -p forge_repo

Notes

• The model list is curated; it can be updated as OpenAI adds/changes models or as modality metadata becomes reliably available via the API.