Multiple NamedTempFiles are created in the same directory. Therefore,
it happens that the cluster marker test can interfere with each other.
This commit separates the cluster marker tests by creating a temporary
directory in which the cluster marker file gets created.

This fixes #1626.

…icate is true

We use StyledTable::add_kv_row_if to check if a value exists. We should only access the
value if the predicate passes. That's why this command changes the value parameter into
a closure that is evaluated if the predicate is true.

This fixes #1631.

To test it: `cargo xtask generate-table-docs`

Unrelated fix to some dependencies that broke my build locally.

Use Rust doc for documenting DF table schemas

This commit changes the place where the table docs need to be registered
to be included in the automatic table docs generation to a single place.

This commit updates the sql table column descriptions wrt feedback
from restatedev/documentation#422.

WipeMode::All will now delete the whole node base dir to also
remove a cluster marker file.

This fixes #1629.

This commit changes the shuffle to always use get_next_message to read
the next outbox message. Before we were using get_message which was only
looking at a specific outbox entry. If this outbox entry was empty, then
the shuffle assumed that the outbox is empty. This did not work if the
outbox contained holes. Now with get_next_message, we always scan until
the next outbox message.

The change itself is trivial. In order to ensure that the change works,
this commit added a few unit tests to ensure the desired behaviour.

Note: With replacing get_message with get_next_message, the shuffle will
perform for every read a RocksDB scan operation. This is highly inefficient
and we should replace this logic with a tailing iterator.

This fixes #1639.

This adds a new configuration key `in-memory-queue-length-limit` under `[worker.invoker]` to enable configuring the segment queue from the configuration file. The default value matches the currently hard coded value.

* Support http1 for request-response protocol deployments

We already have the plumbing to set the request version appropriately based on discovered version (discovery goes through HTTP1.1 by default). We just end up ignoring that version at the last minute by setting the client to be http2 only. With this change, we will discover with http2, and fallback to http1.1 if it doesn't appear to be supported. The fallback request will respect ALPN, so it could technically still end up using http2 against HTTPS endpoints, but this would be odd.

* Accept a flag to force the use of http2

* Add flag to CLI

* Add comments