[moetemp]

What changed?

This PR adds readers. A reader keeps the state of one subscribe, so lang only decodes each item.

• OpenReader checks the options and keeps them. ReadNext reads pages until one has items, the read ends or max_wait passes, and returns that page. Each item carries the resume point after it, as Subscription.next_offset moves per item.
• A Superseded item lists the offsets of the Records that this read gave and the item withdraws. Client Records aren't kept, since nothing withdraws them.
• The reader keeps the create wait, which starts at the first ReadNext, the end of a read with no wait, the newest-last first page, and the pause after a quick empty page. last takes no offset and no END.
• A dropped ReadNext leaves the reader as it was, so a cancel loses no item. A wait of more than a year counts as a year.
• The static Withdraws applies the withdrawal rule to a Producer of an earlier read. epoch::withdraws_producer is that rule for a Producer, and withdraws calls it.
• The rule is K7's. sdk-python's withdraws also has a fallback for a completed task with attempt 0, which compares started times. It isn't here, since Core sends the attempt with C1's field 6. It matters only when Core doesn't know the attempt.

Depends on #223 (K13). Review lane is the Rust reviewers.
Part of the Brook review series.

Why?

sdk-python's _reader.py and _items.py hold the page loop, the waits and the withdrawal rule today. They are the same in every language.

How did you test it?

cargo fmt, cargo clippy -- -D warnings (with and without features), cargo test -p temporalio-streams --all-features and cargo doc pass at this commit. The reader suite runs on the memory Store here, and on Redis in R4 (temporal-community/sdk-streams#4).

[moetemp]

Parked. Nothing calls these runtime methods yet, since sdk-python does this work in Python. The branch stays, and this layer comes back with its first caller, for example a TypeScript or .NET port or Rust read handles.