🚨 [v5] Remove relative position embeddings (for bert like models)

[vasqu]

These embedding types are barely used and make the modeling files just more complex without justifying their existence. Position embedding types still exist in a few models; this PR just addresses the relative_key(_query) ones.

Some stats:

• None of the slow tests use them except bert
• The respective models in those tests together have less than 2k downloads in the last month

cc @hmellor this should remove any clashes with the kwargs you encountered in vLLM :D

[vasqu]

run-slow: flava, instructblib, mra

[vasqu]

This is mostly due to me forgetting to update them in my bert refactor PR --> big diff because the whole refactor is included (same for the roberta example)

[github-actions]

This comment contains run-slow, running the specified jobs:

models: ['models/flava', 'models/mra']
quantizations: [] ...

[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]

run-slow: instructblip

[github-actions]

This comment contains run-slow, running the specified jobs:

models: ['models/instructblip']
quantizations: [] ...

[github-actions]

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

run-slow: albert, align, altclip, bert, bert_generation, big_bird, blip, blip_2, bridgetower, bros, camembert, canine, chinese_clip, clap, data2vec, dpr

[vasqu]

run-slow: bert, roberta, albert, mra, instructblip, blip_2, flava

[github-actions]

This comment contains run-slow, running the specified jobs:

models: ['models/albert', 'models/bert', 'models/blip_2', 'models/flava', 'models/instructblip', 'models/mra', 'models/roberta']
quantizations: [] ...

[vasqu]

Failing slow tests are the same as in main 👀