[docs] Describe how searches by timestamp read offloaded ledgers

[KannarFr]

Completes the "Read performance for object storage" section of the tiered storage overview with the search by timestamp in offloaded ledgers, following apache/pulsar#26778 and apache/pulsar#26782.

Three paragraphs are added after the existing one, which is unchanged:

• How a seek by timestamp, a cursor reset to a timestamp and message TTL checks read a ledger offloaded with the jclouds offloader, and what changes in Pulsar 5.0.0 and later: the search reads the first entry of a data block when it is close to the midpoint of its range.
• The offsets that the jclouds offloader remembers while skipping entries: the pulsar.jclouds.readhandleimpl.learnedoffset.interval.bytes JVM property (default, minimum, how to disable it), how it combines with the backward search already described, what it retains and for how long, and its scope.
• What to inspect before tuning: the message find metrics documented in [docs] Document the OpenTelemetry metrics of the message position find by timestamp #1236 and the offload metrics, and the effect of the read buffer and block sizes.

The measured numbers are left out of this page, like for the other settings it describes. For reference, with the default 64 MiB data blocks a cold search goes from 156 to 44 ranged reads on average (details in apache/pulsar#26778 and apache/pulsar#26782).

The changes ship in Pulsar 5.0, so the same text is added to docs/ and to versioned_docs/version-5.0.x/.

✅ Contribution Checklist

• I read the contribution guide
• I updated the versioned docs

🤖 Generated with Claude Code

[lhotari]

LGTM. Thanks!