0.26.0

BREAKING CHANGES

• A batch of one reply read as a sequence is now the batch, not the reply.
  Vec<T> and tuples over a pipeline or a transaction retaining exactly one
  command used to read that command's own reply; they now read a one-element
  batch, as they already did above one retained command. A caller who queued a
  single collection-returning command -- pipeline.lrange(..).queue() read as
  Vec<String>, or a BLPOP read as (String, String) -- now asks for
  Vec<Vec<String>> and ((String, String),), or keeps the old reading by
  naming the reply's own type. Every non-sequence form is unchanged.

Fixed

• A pipeline or a transaction retaining exactly one reply reads as a collection.
  Pipeline::execute::<Vec<T>>() failed with CannotParseSequence when the
  forget flags left a single reply, while the same code with two surviving
  replies succeeded -- so a caller looping over a variable number of commands had
  to special-case the count. A batch of one carries two readings, the sequence of
  one and the reply itself, and the batch deserializer resolved them by picking
  the reply unconditionally. It now resolves them the way serde already tells them
  apart: the sequence forms answer for the batch, every other form is answered by
  the lone reply. Transaction::execute carried the same shortcut and now goes
  through the very same deserializer, so the two agree on every type. The same
  defect also made (String,) unreadable on a one-command pipeline, which is
  what examples/pipelining.rs shows; it now works.

Internal

• EXEC's reply is read by the pipeline's batch deserializer. The transaction
  had its own DeserializeSeed and SeqAccess deserializer duplicating what
  RespBatchDeserializer does; EXEC's elements are now handed out as responses
  of their own -- a refcount bump each, no byte copied -- and read as a batch.
  A forgotten reply is dropped unread there as it already was in a pipeline,
  errors included.

• into_collection_iter states its bounds without an #[expect]. The head
  increment is a saturating_add, so no arithmetic_side_effects warning is
  raised for the expectation to catch and the reasoning stays where the index is
  computed.

• tape_memory reads its nodes with as_chunks. chunks_exact with a
  constant size is what clippy::chunks_exact_to_as_chunks names; the array
  chunks it yields also drop the try_into().unwrap() each word went through.