[mem_cache][12/N] refactor: stop allocators round-tripping through the KV pool

[alphabetc1]

Motivation

Part of #25371. Rebased onto current main and renumbered -- the branch had been sitting
since June and [6/N] had meanwhile been taken by #30249.

Allocators reach through the KV pool for two things they do not need it for:

1. get_cpu_copy / load_cpu_copy pass-throughs. TokenToKVPoolAllocator and
   SWATokenToKVPoolAllocator each define these purely to forward to self._kvcache,
   and BaseTokenToKVPoolAllocator declares them only to raise NotImplementedError.
   Retraction backup is a pool operation; routing it through the allocator adds a layer
   that carries no decision.
2. Round-trips through the pool to read a tensor the allocator itself owns. The
   allocator allocates full_to_swa_index_mapping / full_to_hisparse_device_index_mapping
   in __init__, hands the same object to the pool via register_mapping, and writes
   it on every alloc and free. self._kvcache.translate_loc_from_full_to_swa(x) is then
   literally self.full_to_swa_index_mapping[x] with an extra call frame.

Modifications

• Drop get_cpu_copy / load_cpu_copy from allocator/base.py, allocator/token.py,
  allocator/paged.py, allocator/swa.py. The two call sites in Req.save_kv_cache /
  Req.load_kv_cache go through token_to_kv_pool_allocator.get_kvcache() instead.
• SWATokenToKVPoolAllocator.translate_loc_from_full_to_swa and the HiSparse allocators'
  get_last_loc_hisparse_device / free_hisparse / free_compressed index the
  allocator-owned mapping directly.
• full -> compressed is deliberately left as a pool call: it is genuine C4-layout
  arithmetic ((i + 1) % compress_ratio), a pool responsibility, not a mapping lookup.
  Comments at both sites say so, to keep the next reader from "simplifying" it too.

No behavior change: every removed indirection resolved to the same tensor index or the
same pool method.

Accuracy Test

Rebase resolution, verified on an H200 devbox against a pristine main baseline on the
same box:

      rebased branch:   1746 passed, 1074 skipped, 201 subtests passed
  pristine main:        1746 passed, 1074 skipped, 201 subtests passed

(test/registered/unit/mem_cache/ plus spec/test_resolve_swa_kv_pool.py; --ignore on
test_umbp_store.py, which needs the mori package that is absent from the image.) Zero
delta, so the 1074 skips are pre-existing.

The rebase kept the original diff shape exactly (+32/-45 across 7 files). Three of the
four conflicts were context drift around the deleted methods -- main had added
translate_kv_indices_for_transfer to allocator/base.py, _copy_for_free_group to
allocator/token.py, and PureSWATokenToKVPoolAllocator plus swa_free_group to
allocator/swa.py. All are preserved; PureSWATokenToKVPoolAllocator is byte-identical
to main, and it overrides translate_loc_from_full_to_swa with an identity, so it is
unaffected by the change here. The fourth conflict was real: main had reworked
retraction backup onto RetractionBackup(cpu_tensors=...), so the resolution keeps that
structure and only reroutes the call through get_kvcache().

The ownership assumption was re-checked against current main rather than trusting the
June-era diff: allocator/swa.py builds the tensor, register_mapping hands the same
object to the pool, and the pool's translate_loc_from_full_to_swa indexes that very
tensor -- so the direct index is identical, including the -1 sentinel row. A repo-wide
grep confirms no allocator-level get_cpu_copy / load_cpu_copy caller or override
survives; every remaining one is on a pool.

Benchmark and Profiling Results

Not applicable -- removes call frames, changes no kernel or memory layout.

Checklist

• Format your code according to the Code Formatting with Pre-Commit.
• Add unit tests as outlined in the Running Unit Tests.
• Update documentation / docstrings / example tutorials as needed, according to Writing Documentation.
• Provide throughput / latency benchmark results and accuracy evaluation results as needed, according to Benchmark and Profiling.
• For reviewers: If you haven't made any contributions to this PR and are only assisting with merging the main branch, please remove yourself as a co-author when merging the PR.
• Please feel free to join our Slack channel at https://slack.sglang.ai to discuss your PR.

🤖 Generated with Claude Code

---

CI States

Latest PR Test (Base): ❌ Run #32693512747
Latest PR Test (Extra): ❌ Run #32693512660
Latest PR Test (AMD ROCm 7.2): ❌ Run #32693512729

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!