Fix the eval_use_gather_object flag usage

[ducha-aiki]

What does this PR do?

Fixes #36213

After the first eval, the 2nd eval is crashing, while trying to concat batches with different shapes, as if the flag eval_use_gather_object stopped working. In evaluation_loop, the self.gather_function is reset, and the eval_use_gather_object is not used

This PR fixes that, using the same code line, which is used at accelerator creation.

@muellerzr and @SunMarc

[ducha-aiki]

The test failure seems to be unrelated:

=================================== FAILURES ===================================
______________ LlavaForConditionalGenerationModelTest.test_config ______________
[gw3] linux -- Python 3.9.21 /usr/local/bin/python3

self = <tests.models.llava.test_modeling_llava.LlavaForConditionalGenerationModelTest testMethod=test_config>

    def test_config(self):
>       self.config_tester.run_common_tests()

tests/models/llava/test_modeling_llava.py:201: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
tests/test_configuration_common.py:205: in run_common_tests
    self.check_config_can_be_init_without_params()
tests/test_configuration_common.py:169: in check_config_can_be_init_without_params
    config = self.config_class()
E   AssertionError: ValueError not raised

[SunMarc]

Thanks for the PR ! LGTM ! Could you try to craft a test that fails before this PR ? Maybe you can take inspiration from test_eval_use_gather_object in test_trainer.py ?