[GPU] fix sdpa_micro_gqa_single_token for multi batch (#34192)
### Details:
- This PR fixed `sdpa_micro_gqa_single_token` to work correctly for
multi batch.
- updated to create an internal buffer for `sdpa_micro` for all stages
including `generate` phase.
- fixed the offset calculation logic for input and output in
`sdpa_micro`.
### Tickets:
- 181161
