fix(interpretability): allow inf-masking SHAP on the remote backends

[safaricd]

Changes

• The guard refused every remote backend. _model_has_inf_passthrough read only get_inference_config(), which no client estimator has, so get_tabpfn_inf_explainer rejected configurations that work. The premise that PASSTHROUGH_INF is local-only is wrong: inference_config is a field both the managed API and a self-hosted container accept, and each forwards it to the TabPFN it runs.
• It now reads either place the flag can live — a local model's resolved config as before, or the raw inference_config argument otherwise. A local model's resolved config still wins, which a test pins.
• Both backends were tested, not assumed. +inf without the flag is rejected either way (container 500, API 422 Input X contains infinity); with it both predict normally, and the API's output moves away from the finite baseline, so the masking reaches the model.
• Docs and tests corrected — the path was documented as local-only.
• Self-hosted JSON payloads additionally need the hosted estimator's non-finite serialization fix, without which +inf never leaves the client. The managed backend uploads Parquet and is unaffected.

Reviewer notes

• Depends on feat(hosted): add use_kv_cache, fix estimator type and non-finite JSON tabpfn-client#372 for the self-hosted path; merge that first. The managed API path works against released tabpfn-client.
• New tests use stubs, since tabpfn-client is not a test dependency here. They are currently the only tests in this file that run without local model weights.
• Suite on this branch: 53 passed vs 49 on main, with the same 51 pre-existing failures (TabPFNLicenseError — no local weights in this environment).

🤖 Generated with Claude Code

https://claude.ai/code/session_01AsHuhQhqWz3EkVVtPbdBBa