[Bugfix] Fix incorrect single-token saves in v1 (LMCache#653)

* [bugfix] fix the concurrency bug in non-blocking put

---------

Co-authored-by: ApostaC <jc4xvyp@outlook.com>

* [XWKV-48] Implement gismo backend (LMCache#7)

* [XWKV-48] Cache fd to optimize performance (LMCache#11)

Cache fd to optimize performance.

Add thread pool to read/write file in parallel.

* [XWKV-48] Support remotely read kv (LMCache#12)

As Gismo supports locally read files
written remotely, we need to change our code
to make sure we can read files not cached.

* [XWKV-68] Use vram API to boost performance (LMCache#13)

As vram read/write API to boost performance.

Use new get/put method to read/write meta file.

* [XWKV-68] Use batched contains API

Use mvfs batched contains API to implement
batch contains interface to
boost performance.

Add retry when get metadata in case lmcache
reads faster than backend.

---------

Co-authored-by: Jinwen <287310886@qq.com>

* [bugfix] fix the concurrency bug in non-blocking put

---------

Co-authored-by: ApostaC <jc4xvyp@outlook.com>

…pter (LMCache#13)

Add sync_mode parameter to LMCacheMPWorkerAdapter so that sync-mode
specific behavior (QUERY_PREFETCH_STATUS_WITH_REQ_ID polling and
blocking retrieve waits) only activates when explicitly enabled.
This prevents the default async path from being affected.

- Add sync_mode: bool = False to LMCacheMPWorkerAdapter.__init__
- Gate prefetch status polling in submit_retrieve_request()
- Gate blocking wait in batched_submit_retrieve_requests()
- Pass sync_mode through create_worker_adapter() and connector init
- Rename LMCacheMPConnector to LMCacheMPConnectorDynamic

Signed-off-by: crclq2018@gmail.com

Signed-off-by: rigginschen <rigginschen@tencent.com>
Co-authored-by: rigginschen <rigginschen@tencent.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>