Described data copy cases (#2698)

Author: olegkkruglov

doc/sources/input-types.rst

@@ -39,3 +39,14 @@ like `Polars <https://pola.rs>`__.
 Extension currently does not offer accelerated routines for input types not listed
 here - when receiving an unsupported class, estimators will fall back to stock Scikit-Learn to
 handle it, so make sure to convert them to a supported type when using Extension.
+
+.. warning::
+  In some cases data passed to estimators might be copied/duplicated during calls to methods such as fit/predict.
+  The affected cases are listed below.
+
+  - Non-contiguous NumPy array - i.e. where strides are wider than one element across both rows and columns
+  - For SciPy CSR matrix / CSR array index array is always copied.
+  - Heterogeneous NumPy array
+  - If SYCL queue is provided for device without `float64` support but data are `float64`, data are copied with reduced precision.
+  - If :ref:`Array API <array_api>` is not enabled then data from GPU devices are always copied to the host device and then result table 
+    (for applicable methods) is copied to the source device.