[Tests] Cast Hubert model tests to fp16

[anton-l]

Problem
Some Hubert integration tests are flaky, throwing GPU OOM on model loads (GPUs are 16GB). CI logs: https://github.com/huggingface/transformers/actions/runs/1274012618

FAILED tests/test_modeling_hubert.py::HubertModelIntegrationTest::test_inference_ctc_batched
FAILED tests/test_modeling_hubert.py::HubertModelIntegrationTest::test_inference_emotion_recognition
FAILED tests/test_modeling_hubert.py::HubertModelIntegrationTest::test_inference_intent_classification
FAILED tests/test_modeling_hubert.py::HubertModelIntegrationTest::test_inference_keyword_spotting
FAILED tests/test_modeling_hubert.py::HubertModelIntegrationTest::test_inference_speaker_identification

Solution
Cast models and reference tensors to fp16.

cc @patrickvonplaten