Transition to default int4 compression configs from optimum-intel

[nikita-savelyevv]

Changes

• For 4BIT_DEFAULT, the compression config is now always imported from optimum-intel
• Remove all default configs from openvino.genai, rely only on default configs in optimum-intel
• Adopt dataset preparation logic from optimum-intel

Breaking changes:

• Old compression dataset specification is deprecated. For example "wikitext,wikitext-2-v1,train[:1000],text" will not work anymore. A string "wikitext2" should be used instead.
• Compression with data-aware methods will produce different results, potentially worse.
• Default int4 configs will no longer be matched by folder names. To match the default config, one of these options must hold:
  • model is exported with python convert.py -m <model_id>
  • model is exported with python convert.py -m <model_path>, where model_path contains a config.json file with a _name_or_path field containing a model id
  • model is exported with python convert.py -m <model_path>, where model_path contains a config.json file with a _name_or_path field containing another model_path_2, such that the condition from the bullet point above holds for it

Related tickets
147470