fix(unified-memory): propagate unified-memory lazy checkpoint policy and handle unallocated session slots

[seokwoosong]

Motivation

With --mamba-radix-cache-strategy extra_buffer_lazy, a request holds one ping-pong checkpoint slot and the second is allocated only when a track boundary needs it. The unified-memory pool factories never forwarded enable_mamba_extra_buffer_lazy, so unified pools allocated both slots up front while the scheduler ran the lazy policy.

Forwarding the flag means unified pools now hold ping-pong buffers whose unallocated entries are -1. free_mamba_cache already filters those, but streaming-session cleanup freed the whole buffer and counted numel() slots. The same gap exists on the static pool in lazy mode.

Modifications

• Forward enable_mamba_extra_buffer_lazy through init_unified_mamba_pools, init_unified_mamba_swa_pools and UnifiedHybridReqToTokenPool, as a required keyword, so a caller that drops it fails with a TypeError instead of silently running eager.
• StreamingSession._free_slot_mamba frees only the allocated (!= -1) ping-pong entries, and session_held_mamba_slots counts only those. This applies to every pool, not just unified ones.
• The shared unified allocator test fixture (python/sglang/test/unified_allocator_fixtures.py) passes the flag explicitly.
• Regression tests: the policy reaches both unified factories through the production configurator, and session cleanup and accounting with -1 entries on unified and static allocators.

Accuracy Tests

Validated on an NVIDIA B200 against main 048b2a07f3, with this PR rebased onto it:

• test_unified_tri_pool.py: 49 tests and 100 subtests pass, including the two new ones (47 on main). The related unified-memory, Mamba and streaming-session test files give the same results as main.
• Inkling (thinkingmachines/Inkling, revision test) with --enable-unified-memory --mamba-radix-cache-strategy extra_buffer_lazy and --enable-deterministic-inference, 32 LongBench prompts:
  • main: the scheduler aborts at its first idle leak check with pool memory leak detected!, the Mamba pool counting 32 slots more than it holds, one per request;
  • with this change (run together with [Bugfix] Fix unified-memory compaction gates and pending page reuse #39982 and fix(inkling): translate unified-memory checkpoint destinations to physical slots #38229): logprob match, prefill cache hit and decode cache hit are all at avg_kl_div=0.0.

After rebasing onto main 6cec8f98ce (which includes #39982), the tests that build pools through the shared fixture pass: test_unified_pending_event_batches.py, test_unified_dynamic_gate_capacity.py and test_unified_float_move_gate.py, along with test_unified_tri_pool.py (49 tests).

Speed Tests and Profiling

No speed impact expected; the session count adds one device read per held slot, and it runs only in the idle leak check.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations. Not applicable: no new public option.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

---

CI States

Latest PR Test (Base): 🚫 Run #37193489703
Latest PR Test (Extra): ❌ Run #37193489469
Latest PR Test (AMD ROCm 10): ❌ Run #37193489678