feat: expose server-reported fit and predict timings

[ggprior]

Adds server-reported timings to TabPFNClassifier and TabPFNRegressor, so users can tell time spent waiting for the server to start the work apart from time spent doing it.

model.fit(X_train, y_train)
model.predict(X_test)
model.get_timings()
# {"fit": {"elapsed_s", "queue_wait_s", "train_set_transform_s", "fit_s"},
#  "predict": {"test_set_transform_queue_wait_s", "test_set_transform_s",
#              "predict_queue_wait_s", "predict_s"}}

• The estimators keep the values as fit_timings_ (set by fit()) and last_predict_timings. get_timings() returns both.
• PredictionResult gains timings. ServiceClient.fit_with_result() returns a FitResult with the fitted id and the timings; ServiceClient.fit() still returns only the id.
• A fit that is still pending after /fit takes its timings from the final status poll.
• A regressor prediction split into several requests (output_type="full" above the row limit) reports the per-stage sum.
• api_models.py gains FitTimings, PredictTimings and the optional timings fields, copied from the API schema.

Compatibility

• Against a server that does not report timings, every value is None; nothing else changes.
• Predict timings live in a private attribute behind a read-only property, because check_estimator requires predict() to leave the estimator's __dict__ unchanged.
• Tests that patched InferenceClient.fit now patch InferenceClient.fit_with_result, which the estimators call.

🤖 Generated with Claude Code

https://claude.ai/code/session_01A99qZsvdCamLPCMD77vkDw