Use native iris functions in multi-model statistics (#1150)

* Branch off lazy mmstats PR

* Remove code related to lazy evaluation

* Turn off tests for lazy data

* Enforce dtype float32 in test cubes

* Fix dtype issue

* Better note on dtype fix

* improve warning about std_dev

* Catch errors in cube alignment with nice error message

* Catch concat errors in compute_eager

* Remove doc statement about lazy evaluation

* Regenerate sample data for failing regression tests

esmvalcore/_recipe_checks.py

@@ -9,7 +9,8 @@
 import yamale
 
 from ._data_finder import get_start_end_year
-from .preprocessor import PreprocessingTask, TIME_PREPROCESSORS
+from .preprocessor import TIME_PREPROCESSORS, PreprocessingTask
+from .preprocessor._multimodel import STATISTIC_MAPPING
 
 logger = logging.getLogger(__name__)
 
@@ -191,7 +192,7 @@ def extract_shape(settings):
 
 def valid_multimodel_statistic(statistic):
     """Check that `statistic` is a valid argument for multimodel stats."""
-    valid_names = ["mean", "median", "std", "min", "max"]
+    valid_names = ['std'] + list(STATISTIC_MAPPING.keys())
     valid_patterns = [r"^(p\d{1,2})(\.\d*)?$"]
     if not (statistic in valid_names
             or re.match(r'|'.join(valid_patterns), statistic)):