Whisper pipeline: add perf metrics

[as-suvorov]

This PR adds:

• support perf metrics

Common Todos for Whisper support:

• Long-form audio support with parallel chunking.
• update documentation
• add cpp, python samples tests
• support timestamps streaming
• expose only meaningful parameters in GenerationConfig (task, language, return_timestamps, etc)
• Move all whisper pipeline files to dedicated subfolder
• Whisper pipeline doesn't need tokenizer, it uses detokenizer only. Implement detokenizer only initialization for ov::genai::Tokenizer
• Check discrete GPU. Integrated GPU works as expected.
• Investigate use of RemoteTensor for GPU
• Add batch
• Add sampler, inherit WhisperGenerationConfig from GenerationConfig
• Investigate language autodetection with single decoder (without past) call
• Update python bindings cmake to include whole directory instead of explicit list of files
• Add samples with audio preparation examples
• Add links to audio files so users can download them in samples
• Move supported models list from samples README to common supported models section
• Avoid building GenAI in each tests job as it takes a lot of time
• Double check FP32 support
• Fix tests sporadic fails. Sometimes whisper model cannot be downloaded from HF due to network issues
• Fix stop criteria. Current approach stops on eos_token which is no speech token. But there could be more speech tokens further which are wrongly skipped now
• Fix distil whisper accuracy, match with HF
• Fix en models accuracy with timestamps, match with HF
• Try to trim input_ids cache between chunks for long-form audio to match HF

Completed:

• support different languages, language autodetection
• support translation
• support timestamps
• Long-form audio support with sequential chunking.

Current limitations:

• No resampling during preprocessing. Input raw speech should have 16k Hz sampling rate
• No normalization during preprocessing. Input raw speech should be normalized to near [-1, 1] range

Tickets: CVS-147994, CVS-146010, CVS-152523