Test all models

[noahho]

Motivation and Context

This patch expands our test coverage to run against all available TabPFN models while keeping CI time reasonable.

• Introduce ModelSource in tests to programmatically fetch model filenames.
• Restructure parameter grids so we fully exercise one “primary” model (broad grid) and run a fast “smoke” sweep across the remaining models.
• Reduce reliance on hardcoded "auto" model selection inside tests to make failures attributable to specific model artifacts.

This addresses gaps identified in “Run tests with all models” and aligns with feedback that initialization and interface tests should validate each shipped model.

---

Public API Changes

• No Public API changes
  (Only test code is modified.)

---

How Has This Been Tested?

• Confirmed the new grids keep test duration practical on CPU-only environments.

---

Checklist

• The changes have been tested locally.
• Documentation updated: N/A (tests only).
• CHANGELOG.md entry: N/A (tests only).
• Code follows project style (ruff format & ruff check).
• Considered impact on public API: None (tests only).

---

Implementation Notes

• Built:
  • _full_grid: exhaustive combos only for the first model path.
  • _smoke_grid: a single, fast combo for each remaining model path.
• Combined as all_combinations = list(_full_grid) + list(_smoke_grid).

---

Prior Work / Acknowledgements

Big thanks to the work in PR #437 “Run tests with all models” (by @martino-vic) which motivated this direction and highlighted the need to validate every shipped model. That PR is good prior art but currently not mergeable (pending checks/CLA and minor integration issues). This patch folds in the workable pieces (programmatic model discovery + targeted grid split) and aligns them with the current test suite and style rules.

[noahho]

/gemini review

[gemini-code-assist]

Code Review

This pull request significantly improves test coverage by parameterizing tests to run against all available TabPFN models, which is a great enhancement. The approach of using a full grid for a primary model and a smoke test for others is clever and keeps CI times reasonable. I've identified a minor robustness issue where the tests could crash if no models are found. Adding a check to gracefully skip the tests in this scenario would make the test suite more resilient. Overall, this is a solid contribution to improving the project's test quality.

[LeoGrin]

bugbot run

[LeoGrin]

LGTM, just two comments

[LeoGrin]

not sure about the failing test, probably just easier to go over the precision limit by random change if we test more models right?