Enhance several unit tests (#62477) (#62776)

* Manually release predictor_tuned
* Add indices to no_cast_list to keep it as fp32
* Set both atol and rtol for the fp16 test_trt_convert_solve
* Merge branch 'rewang/fix_test_sparse_fused_attention_seed' into 'nv-2.6.0'

Fix test_sparse_fused_attention random seed

See merge request dl/paddle/paddle!312

---------

Signed-off-by: rewang <rewang@nvidia.com>
Co-authored-by: Ryan Jeng <rjeng@nvidia.com>

Author: zlsh80826

?test/cpp/inference/api/trt_dynamic_shape_test.cc

@@ -191,6 +191,7 @@ void TestTunedDynamic() {
     output_t->copy_to_cpu(out_data.data());
   };
   check_func(predictor_tuned.get());
+  predictor_tuned.reset(nullptr);
 
   // check tuned_dynamic_shape
   AnalysisConfig config;

?test/ir/inference/test_trt_convert_lookup_table.py

@@ -80,6 +80,7 @@ def generate_input2(dims, attrs: List[Dict[str, Any]]):
                     )
                 },
                 outputs=["out_data"],
+                no_cast_list=["indices"],
             )
 
             yield program_config

?test/ir/inference/test_trt_convert_solve.py

@@ -90,7 +90,7 @@ def clear_dynamic_shape():
         yield self.create_inference_config(), (1, 3), 1e-5
 
         self.trt_param.precision = paddle_infer.PrecisionType.Half
-        yield self.create_inference_config(), (1, 3), 1e-3
+        yield self.create_inference_config(), (1, 3), (1e-3, 1e-3)
 
     def test(self):
         self.run_test()

?test/legacy_test/test_sparse_fused_attention_op.py

@@ -42,6 +42,7 @@ def get_cuda_version():
 )
 class TestSparseAttentionAPI1(unittest.TestCase):
     def setUp(self):
+        paddle.seed(0)
         self.batch_size = 16
         self.num_heads = 16
         self.seq_len = 128
@@ -134,6 +135,7 @@ def test_dygraph(self):
 
 class TestSparseAttentionAPI2(TestSparseAttentionAPI1):
     def setUp(self):
+        super().setUp()
         self.batch_size = 16
         self.num_heads = 16
         self.seq_len = 128
@@ -144,6 +146,7 @@ def setUp(self):
 
 class TestSparseAttentionAPI3(TestSparseAttentionAPI1):
     def setUp(self):
+        super().setUp()
         self.batch_size = 16
         self.num_heads = 16
         self.seq_len = 512
@@ -154,6 +157,7 @@ def setUp(self):
 
 class TestSparseAttentionAPI4(TestSparseAttentionAPI1):
     def setUp(self):
+        super().setUp()
         self.batch_size = 16
         self.num_heads = 16
         self.seq_len = 512
@@ -164,6 +168,7 @@ def setUp(self):
 
 class TestSparseAttentionAPI5(TestSparseAttentionAPI1):
     def setUp(self):
+        super().setUp()
         self.batch_size = 16
         self.num_heads = 16
         self.seq_len = 512