Refactor KV Cache Manager with a new torch exportable KVCache backend

[geoffreyQiu]

Description

Refactor KV Cache Manager with a new torch exportable KVCache backend

• Introduce KVCacheBackend, split the default Python backend from the public KVCacheManager facade
• Implement KVCacheManagerContext - creation, registry, reference.
• Implement the exportable lookup_kvcache and allocate_kvcache.
• Implement the exportable onboard_kvcache_launch and onboard_kvcachce_wait.
• Implement the exportable offload_kvcache_launch and offload_kvcache_reap_completed (replacing offload_try_wait).

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.