Allow Attention and Experts to be used as standalone modules

[Cyrilvallez]

What does this PR do?

As per the title. Also fix a vllm regresion on experts. Allow Experts and Attention module to be used on their own, such as

import torch
from transformers import AutoConfig
from transformers.models.mixtral.modeling_mixtral import MixtralExperts
config = AutoConfig.from_pretrained("mistralai/mixtral-8x7b-v0.1")
experts = MixtralExperts(config)
experts.forward(torch.randn((1, 64, experts.config.hidden_size)))

 File "/home/harry/transformers/src/transformers/integrations/moe.py", line 317, in forward
    experts_forward = ALL_EXPERTS_FUNCTIONS[self.config._experts_implementation]
                      ~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/harry/transformers/src/transformers/utils/generic.py", line 1070, in __getitem__
    return self._global_mapping[key]
           ~~~~~~~~~~~~~~~~~~~~^^^^^
KeyError: None

[vasqu]

Could we add a small test so that we don't have a regression?

[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.

[vasqu]

LGTM, just 2 nits

[github-actions]

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

run-slow: afmoe, aimv2, albert, align, altclip, apertus, arcee, aria, audio_spectrogram_transformer, audioflamingo3, bamba, bart, bert, bert_generation, bigbird_pegasus, biogpt