ext test case (DeepLink-org#721)

* set flash-attention lib default path depends on python version.

* rms_norm

* rotary_emb

* apply_penalty

* destindex_copy_kv

* token_attention

* token_softmax_reducev

* context_attention

* start_loc start from not zero

* mha add GLOBAL_STATE

* multihead_attention_varlen


---------

Co-authored-by: gqw <[email protected]>
Co-authored-by: Gong-air <[email protected]>
Co-authored-by: wey-code <[email protected]>

Author: 3 people