docs(offload): clarify LMCache MP lookup timeout scope

[kvnloo]

Summary

Docs-only follow-up stacked on #2250's current head 94c07412.

The #2250 owner already integrated the separately reported --eviction-policy LRU server-recipe correction. This PR therefore contains only one remaining clarification: document the timeout scopes already implemented by native MP.

• lmcache.mp.mq_timeout is passed to the LMCache scheduler/worker adapters for message-queue operations.
• lmcache.mp.lookup_timeout starts after lookup submission returns and bounds ATOM's polling loop while the tier keeps returning no answer.
• lmcache.mp.lookup_poll_interval controls the sleep between polls.

So lookup_timeout is not a hard wall-clock bound on a blocking adapter submission or status call.

No topology, lifetime, scheduler, runtime-default, server recipe, or LMCache dependency change.

Evidence

CPU characterization run exercised the real ATOM lookup facade with a deterministic adapter/clock alongside the existing MP tests:

• 65 MP tests passed, no failures/skips.
• A submission consuming 8 fake seconds followed by a 3-second polling budget does not time out until the post-submission polling deadline.
• A status call consuming 8 fake seconds and then returning a hit is accepted before the outer polling deadline is checked again.

I refreshed the source against #2250 head 94c07412; the lookup control flow remains unchanged for this contract.

The branch is merged forward with owner changes rather than force-pushed. Net diff against #2250 remains one README file, +8/−0.

AI-assisted source review and drafting; no server/GPU timeout behavior claimed.