Ignore converted datasets in invalid input states

Fixes a bug reported by @jennaj where using an input dataset from a history
import (which contains implicitly converted datasets in discarded state)
in an input that would require that exact same implicit conversion
would fail with `Job output deleted by user before job completed`.

This will make the whole implicit coversion system much more robust,
since we didn't even exclude datasets in error state, so as soon as one
conversion failed ... you'd be stuck.

lib/galaxy/model/__init__.py

@@ -3945,7 +3945,9 @@ class Dataset(Base, StorableObject, Serializable):
 
     non_ready_states = (states.NEW, states.UPLOAD, states.QUEUED, states.RUNNING, states.SETTING_METADATA)
     ready_states = tuple(set(states.__members__.values()) - set(non_ready_states))
-    valid_input_states = tuple(set(states.__members__.values()) - {states.ERROR, states.DISCARDED})
+    valid_input_states = tuple(
+        set(states.__members__.values()) - {states.ERROR, states.DISCARDED, states.FAILED_METADATA}
+    )
     no_data_states = (states.PAUSED, states.DEFERRED, states.DISCARDED, *non_ready_states)
     terminal_states = (
         states.OK,
@@ -4691,7 +4693,7 @@ def display_info(self):
 
     def get_converted_files_by_type(self, file_type):
         for assoc in self.implicitly_converted_datasets:
-            if not assoc.deleted and assoc.type == file_type:
+            if not assoc.deleted and assoc.type == file_type and assoc.state in Dataset.valid_input_states:
                 if assoc.dataset:
                     return assoc.dataset
                 return assoc.dataset_ldda