[mem_cache][11/N] refactor: extract KVCache and BaseSWAKVPool into pool/base.py

[alphabetc1]

Motivation

Part of #25371. Mechanical Move. First PR of the mem_cache/pool/ layer.

pool/ is the last of the three layer packages and the only one still getting worse.
Since the restructure was agreed, memory_pool.py went from 2257 to 5042 lines and from
11 to 17 classes, because there was nowhere else for a device pool to go. This creates
the package and moves the base layer into it, so every later family move has a target.

Modifications

New mem_cache/pool/base.py, holding only what every device pool derives from:

Symbol	From
KVWriteLoc, unwrap_write_loc	memory_pool.py
KvBufferDesc	memory_pool.py
KVCache (ABC)	memory_pool.py
BaseSWAKVPool (ABC)	base_swa_memory_pool.py (deleted)
GB	memory_pool.py

base_swa_memory_pool.py is deleted: it was a 29-line file holding one cross-family
ABC, sitting under the SWA feature's name while SWAKVPool, DeepSeekV4TokenToKVPool
and the disagg paths all depend on it. It belongs next to KVCache.

No re-export shim: all 44 call sites are updated in this PR, matching how allocator.py
was deleted outright rather than left forwarding.

RadixAttention moves under TYPE_CHECKING in the new module. It is only ever an
annotation (layer: RadixAttention), and the file has from __future__ import annotations; keeping it at runtime made pool/base.py -> layers.radix_attention ->
... -> forward_context -> pool/base.py a genuine import cycle, because pool/base.py
is imported far earlier in the chain than memory_pool.py was. This mirrors how
LayerDoneCounter was already handled.

Two now-unused imports (abc, KVWriteLoc) drop out of memory_pool.py. Both remaining
KVWriteLoc mentions there are in comments.

Accuracy Test

Mechanical move, so the bar is byte-level equality:

• The KVWriteLoc / unwrap_write_loc / KvBufferDesc / KVCache block is
  byte-identical to the original memory_pool.py region.
• The BaseSWAKVPool class body is byte-identical to base_swa_memory_pool.py.
• Only the new module's import header is hand-written.

Verified on an H200 devbox (PYTHONPATH shadowing the image's copy):

              PR branch:   1746 passed, 1074 skipped, 201 subtests passed
  pristine main baseline:  1746 passed, 1074 skipped, 201 subtests passed

Identical command (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) run against both trees on the same box --
zero delta, so the 1074 skips are pre-existing and none were introduced here. MRO was
asserted intact (KVCache in MHATokenToKVPool.__mro__,
BaseSWAKVPool in SWAKVPool.__mro__), and the re-run after the ruff import cleanup gave
the same counts.

Benchmark and Profiling Results

Not applicable -- no runtime behavior changes.

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.

---

Independent of the other in-flight #25371 PRs -- touches no file that #35306, #35643 or
#35644 touch. Once this lands, #35638 can enable its pool role and move the device
pools from its _SHRINKING_MODULES pin into _LEGACY_HOMES.

🤖 Generated with Claude Code

---

CI States

Latest PR Test (Base): ❌ Run #34434260912
Latest PR Test (Extra): ✅ Run #34439738060
Latest PR Test (AMD ROCm 10): ❌ Run #34434260836

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: e6bbc4efc0

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Update the mocked module tree for the new pool import

When test/manual/minimax_m3/test_npu_memory_pool.py runs _load_npu_memory_pool_module(), it stubs sglang.srt.mem_cache as a plain ModuleType and provides only the old memory_pool.unwrap_write_loc. Executing this newly added import therefore raises ModuleNotFoundError: 'sglang.srt.mem_cache' is not a package before any of the standalone NPU pool tests run. The loader needs to stub sglang.srt.mem_cache.pool.base and its unwrap_write_loc symbol, or otherwise load the real package.

Useful? React with 👍 / 👎.