[generate] shape checks in tests compatible with fixed-length caches (+ some minor fixes) (#35993)

* shape checks compatible with static cache

* add test

* tmp

* manually turn on eager attn when we want to output attn

* typo

* generalize to encoder-decoder models

* force compilation on cpu

* tmp commit

* fix static cache shape checks

* models with odd caches

* fix copies

* shorter cache search loop

* use decoder_past_key_values everywhere

* better test variable names and comments

* signature

* rename _check_outputs into _check_generate_outputs

* add comments

* HybridCache future test note

tests/models/led/test_modeling_led.py

@@ -468,12 +468,12 @@ class LEDModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin,
                 ],
             )
 
-    def _check_encoder_attention_for_generate(self, attentions, batch_size, config, seq_length):
+    def _check_encoder_attention_for_generate(self, attentions, batch_size, config, prompt_length):
         # overwrite because LED does not have (bs, num_heads, seq_len, seq_len) shape
         encoder_expected_shape = (
             batch_size,
             config.num_attention_heads,
-            seq_length,
+            prompt_length,
             self.model_tester.attention_window // 2 * 2 + 1,
         )
         self.assertIsInstance(attentions, tuple)