Feature request - C API - support raw data as input for in-place prediction

In-place prediction is supposed to be fast and efficient according to https://xgboost.readthedocs.io/en/stable/prediction.html#in-place-prediction

[C API](https://xgboost.readthedocs.io/en/stable/c.html#_CPPv425XGBoosterPredictFromDense13BoosterHandlePKcPKc13DMatrixHandlePPK9bst_ulongP9bst_ulongPPKf) only supports JSON string as input for in-place prediction which requires conversion from raw data to char array which later is converted back to raw data. It would be more efficient to support raw data without JSON conversion. In my case, the input is a simple 1D float array.

Start at the XGBoosterPredictFromDense C API entry point and trace how its JSON input is converted before prediction. Confirm the supported raw-data shape and pointer contract, then verify that an in-place prediction accepts a 1D float array without JSON conversion and produces the expected result.

Written by the indexing model from the issue text.