### What this PR does / why we need it?
Refactor V1 Attention for better extensibility (prepared for torchair
attention refactor).
**Main changes:**
- Move different kinds of foward into their method respectively, e.g.,
`_forward_prefill_no_cache()`, `_forward_prefill_cache_hit()`,
`_forward_decode_only()`, `_forward_v1_style()`.
### Does this PR introduce _any_ user-facing change?
No.
- vLLM version: v0.10.0
- vLLM main:
14a5d903ab
Signed-off-by: shen-shanshan <467638484@qq.com>
55d0790597