Add whats-new entry for CF datetime dtype inference speedup

Author: aladinor

doc/whats-new.rst

@@ -104,6 +104,20 @@ Deprecations
 ~~~~~~~~~~~~
 
 
+Performance
+~~~~~~~~~~~
+
+- Skip data reads during CF datetime dtype inference when the output dtype is
+  determinable from metadata alone (``use_cftime`` explicitly set, non-standard
+  calendar, non-nanosecond ``time_unit``, non-chunked data, or when
+  ``valid_max`` / ``time_coverage_end`` / a narrow storage dtype bounds the
+  possible values). This can significantly speed up
+  :py:func:`open_dataset` / :py:func:`open_datatree` on remote-backed stores
+  (zarr on S3, icechunk) with many time-encoded variables. A runtime
+  dtype-mismatch guard on the lazy decode wrapper prevents the rare overflow
+  case from being silently cast by dask.
+  By `Alfonso Ladino <https://github.com/aladinor>`_.
+
 Bug Fixes
 ~~~~~~~~~