[GPU] Fix paged_attention mixed-stage internal buffer layout mismatch (with regression test)

[sungeunk]

Root-cause

• Allocation-time micro/non-micro decision could diverge from execution-time decision, causing mixed-stage buffer count mismatch (expected 7, got 4).

Fix

• Added a regression test for mixed-stage paged attention with token_type_ids to catch internal buffer layout mismatch.
• Fixed paged_attention_opt internal buffer micro-SDPA gating to follow runtime decision consistently.

Tickets:

• 188245

AI Assistance:

• AI assistance used: yes
• AI: fix
• User: code review