Fix fp32_ln for various models

[remi-or]

This PR fixes the test test_flash_attn_2_fp32_ln for several models:

• bark was failing the test because it call _flash_attention_forward directly without checking the queries dtype, and so the test could fail if the dtype was torch.float32. To fix this we re-factored out a code block into a function get_target_dtype that takes care of infering whether to cast the fp32 tesnor to fp16 or bf16, and added a called to it before the call to FA
• same for stablelm
• mllama was failing the test because MllamaTextSelfAttention lacks the is_causalattribute, which was added and set to True (it's a text attention so it's causal, as discussed in Mllama fixes #39182)
• same for kosmos2 but the test still fails for many many other reasons

The list of fixed test is here:

FAILED tests/models/bark/test_modeling_bark.py::BarkSemanticModelTest::test_flash_attn_2_fp32_ln - RuntimeError: FlashAttention only support fp16 and bf16 data type
FAILED tests/models/bark/test_modeling_bark.py::BarkCoarseModelTest::test_flash_attn_2_fp32_ln - RuntimeError: FlashAttention only support fp16 and bf16 data type
FAILED tests/models/mllama/test_modeling_mllama.py::MllamaForCausalLMModelTest::test_flash_attn_2_fp32_ln - AttributeError: 'MllamaTextSelfAttention' object has no attribute 'is_causal'
FAILED tests/models/mllama/test_modeling_mllama.py::MllamaForConditionalGenerationModelTest::test_flash_attn_2_fp32_ln - AttributeError: 'MllamaTextSelfAttention' object has no attribute 'is_causal'
FAILED tests/models/stablelm/test_modeling_stablelm.py::StableLmModelTest::test_flash_attn_2_fp32_ln - RuntimeError: FlashAttention only support fp16 and bf16 data type

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: bark, blt, kosmos2, mllama, stablelm

[ArthurZucker]

LGTM thanks