[mem_cache] Free hybrid SWA pages by one representative per page on page_size > 1

[hnyls2002]

Summary

• Add free_swa_segment(idx, *, start_pos), the SWA-only counterpart of free_segment / free_full_segment, and route every kv-row caller through it; free_swa(idx) stays the set-shaped form (alloc rollback)
• On page_size > 1 the SWA side now frees by host-derived page ids (mapping[idx[::page_size]] // page_size) instead of expanding pages, filtering > 0 and dedup'ing through torch.unique, so a hybrid SWA free no longer synchronizes the scheduler stream at any page size (follows [mem_cache] Make free_swa sync-free on page_size == 1 #36723, which covered page_size == 1)
• Add free_page_ids(page_ids) to the paged and token allocators as the page-native primitive (exact pages, no dedup); free_segment is built on it

Details

• Inside a free group the composite queues the gathered SWA page ids (swa_page_ids_group) and flushes them with one free_page_ids, so the group flush is fixed-shape too
• Callers moved to the segment form: window eviction in free_swa_out_of_window_slots, SWARadixCache node frees (free_swa / free / free_full on page-aligned node values), the unified FreeComponentDeviceSlot action; UnifiedSWATokenToKVPoolAllocator splits its existing start_pos branch into free_swa_segment so both hybrid allocators share the API
• Page-level 1:1 between the full and SWA pools holds on every pairing path (extend / decode alloc drive both pools with one seq_lens, HiCache load-back pairs equal-length page-aligned chunks), so the first token of a page maps into the whole peer page

Tests

• test_swa_unittest.py: page-set equivalence against torch.unique over every tail length with the mapping cleared per page, set_sync_debug_mode("error") on the segment path in and out of a group (CUDA), the _SWA_PEER_MAPPED predicate on page_size == 4, tree eviction never reaching the inner free, and the group-ownership test parametrized over page_size in (1, 4)

Original commits

• 80197780d1b0dff8cc3bd556e790de7e7900653c

---

CI States

Latest PR Test (Base): ✅ Run #34162095517
Latest PR Test (Extra): ✅ Run #34162095388
Latest PR Test (AMD ROCm 7.2): ❌ Run #34162095509