Remove the dead cache_trainset_representation flag

[jmkuebler]

Issue

RES-2743 (base of #1219).

Motivation and Context

cache_trainset_representation has been dead since InferenceEngineCacheKV was removed in #1057. Every architecture's get_architecture accepts it and immediately dels it; the KV cache is owned by the inference engine and selected at call time through forward(). The flag survived only because the ArchitectureModule protocol still declared it, which forced every implementation, every loader, and every test call site to carry it.

This removes it end to end:

• ArchitectureModule.get_architecture(config) no longer takes the kwarg; the four architectures drop the parameter and their "accepted for interface compatibility but ignored" notes.
• load_model_criterion_config, load_model and _build_model lose the parameter.
• The opt-in built-model cache (TABPFN_MODEL_CACHE_SIZE) no longer gates on the flag. The gate assumed a fit_with_cache build mutates the module, which it does not, so that fit mode now uses the cache like every other one.
• initialize_tabpfn_model loses its fit_mode argument, whose only job was computing the flag.
• Tests: every call site, the three fake ArchitectureModules, the "mutating build is never cached" test, and the load_model signature tripwire.

#1219 is stacked on top and shrinks to keying the cache on device and dtype.

fomo-fitting side: PriorLabs/fomo-fitting#3776 drops the parameter from every architecture there. It is a draft until this merges and the tabpfn pin is bumped.

---

Public API Changes

• No Public API changes

• Yes, Public API changes (Details below)

• ArchitectureModule.get_architecture no longer accepts cache_trainset_representation. External architectures implementing the protocol must drop the parameter.

• tabpfn.model_loading.load_model and load_model_criterion_config no longer accept cache_trainset_representation.

• tabpfn.base.initialize_tabpfn_model no longer accepts fit_mode.

• With TABPFN_MODEL_CACHE_SIZE set, fit_mode="fit_with_cache" now uses the built-model cache. The default (cache off) is unchanged.

---

How Has This Been Tested?

• ruff check / format clean.
• Cache off (shipping default): tests/test_classifier_interface.py and tests/test_regressor_interface.py pass on cpu and cuda; the touched unit files (test_model_cache, test_model_loading, test_config, test_inference_config, test_inference, test_architectures/, test_save_load_fitted_model) pass.
• Cache on, the interface suites fail with dtype/device collisions, as they did before this change (see fix: key the built-model cache on device and inference dtype #1219, which fixes that and is stacked on this PR).

---

Checklist

• The changes have been tested locally.
• Documentation has been updated (if the public API or usage changes).
• A changelog entry has been added (see changelog/README.md), or "no changelog needed" label requested.
• The code follows the project's style guidelines.
• I have considered the impact of these changes on the public API.

🤖 Generated with Claude Code

[jmkuebler]

cc @priorphil as you authored #1057 any reason to keep cache_trainset_representation?

[oscarkey]

sweeeet

[github-actions]

Reapproving after a push, as this is a PR from an internal author and was already approved.