audio tester class (#45391)
* audio tester
* tweak check repo for audio tester
* audio -> ALM
* ALMTester: no audio/text defaults; better input prep
* udpate test_sdpa_can_dispatch_composite_models to hanlde ALMs
* propagate to other model classes
* cleaner
* updates
* audio_mask_key + updates
* typo
* simplify granite speech
* nits
* some more cleaning
* add test_mismatching_num_audio_tokens
* add get_placeholder_mask
* specific to musicflamingo
* granite speech fix
* let's factorise alm/vlm testers
* make fix-repo
* unskip test_sdpa_can_dispatch_on_flash on qwen2_audio
* should not be skipped
* make fix-repo
* test_mismatching_num_audio_tokens should be skipped for voxtral_realtime
* nit
* _special_token_ids as property and skipped in prepare_config_and_inputs_for_common
* MoE params in common class
* add _TEXT_MODEL_TESTER_DEFAULTS to avoid divergence
* nit
* clearer inits
* _prepare_modality_inputs return dict
* format
* split line for readability
* ran python utils/check_modular_conversion.py --fix_and_overwrite
* testing auto cancel
* testing auto cancel - part 2
* remove comment
* udpate granite speech plus tests
* fix test
---------
Co-authored-by: eustlb <94853470+eustlb@users.noreply.github.com>
