perf: use size with offset in lazy loading BTreeMap items (#337)

This PR improves performance by removing duplicated reads of lazy object
size.

Lazy objects had duplicated reads for the size of the stored data. First
when it lazily stores the object by reference (it needs to know the size
of the data to advance the offset). And then the second time when the
object needs to be downloaded.

Now the size of the data is stored next to the offset on the first read
and skipped reading on the second one.

Google sheets [CSV
reports](https://docs.google.com/spreadsheets/d/1uHaC0RJ_v9XG8fsP_k236zlda5eFaufpUKAwqxJsd5M/edit?usp=sharing).

btreemap regressed: 11, instructions Δ%: max +3.81% 🔴
btreemap improved: 70, instructions Δ%: min -5.42%, median -1.15% 🟢

Author: maksymar