Feat: Support Whisper + FlashAttention2

[hongziqi]

Is your feature request related to a problem? Please describe.
Currently, the Whisper models integrated in mindnlp do not leverage optimized attention backends like FlashAttention2, which are widely adopted in large-scale Transformer architectures for both speed and memory efficiency.

Describe the solution you'd like
We would like to introduce FlashAttention2 support for Whisper models in mindnlp, leveraging MindSpore’s built-in operations and adapting relevant components for compatibility. Specifically, the solution includes:

1. Replace flash-attn package functions with MindSpore APIs:
   Use mindspore.ops.flash_attention_score as a substitute for the original flash_attn_func and flash_attn_varlen_func provided by the flash-attn library, enabling FlashAttention functionality directly within the MindSpore framework.

2. Re-implement bert_padding utilities in MindSpore:
   Convert the bert_padding utility functions used in flash-attn (such as index_first_axis and index_put_first_axis) into equivalent implementations based on MindSpore, ensuring correct behavior for batched variable-length input.

3. Add a new utility module model_flash_attention_utils:
   Introduce a new helper module that includes functions like _flash_attention_forward and other FlashAttention2-specific utilities, adapted for Whisper and written using MindSpore APIs.

4. Provide a new model class WhisperFlashAttention2:
   Create a new Whisper variant that integrates the above FlashAttention2 optimizations, preserving the original model architecture while improving runtime efficiency and scalability.