chore(dkg): remove summary's transcripts for remote subnets (2/4) - #11221

…inity#11220)

Since the work performed to speedup the latency of remote DKG requests,
their results are delivered as part of data blocks instead of summaries.
The `transcripts_for_remote_subnets` field in summaries is thus now
unused. This PR begins the migration to remove it entirely. Once it is,
the batch delivery can be rewritten more cleanly as we would know that
summary heights never contain any consensus responses.

We cannot simply replace the field with a `BackwardsCompatible<Vec<_>,
_>` (i.e. replace the `Vec<_>` with an `Option<Vec<_>>`) because the
field is a `repeated`, and an empty vector is not differentiable from an
absent one. In contrast, `Some(vec![])` and `None` have different
hashes. For that reason, we need to add a new sentinel field to the
protobuf definition that will be set to `true` to explicitely indicate
that the field is absent. The entire rollout can be reviewed by looking
at the stacked PRs and is described below (together with justifying why
upgrades & downgrades are safe).

- V0 (current) <-> V1 (after this PR):
- The goal of V1 is to keep the exact same representation (including
hash) of V0's summaries while still understanding the sentinel and
compute the same hash as V2 for V2's summaries.
- Upgrade V0 -> V1: A V0's summary doesn't set the sentinel so the field
deserializes as `Some(vec![])` and the hash of an empty vector is
appropriately computed.
- Rollback V1 -> V0: a V1's summary doesn't set the sentinel just yet.
V0 does not read it anyways and always includes an empty vector in the
hash, same as V1.
- V1 <-> V2 ([dfinity#11221](dfinity#11221)):
- The goal of V2 is to actually stop hashing an empty vector, and
instead just ignore it.
- Upgrade V1 -> V2: V2 still understands `Some(vec![])` so will compute
V1's summary's hash identically. For its own summaries, it starts to
fill the sentinel (because the transcripts are `None`).
- Rollback V2 -> V1: V1 understands the sentinel and thus ignores the
transcripts while computing the hash.
- V2 <-> V3 ([dfinity#11223](dfinity#11223)):
    - The goal of V3 is to remove the field entirely.
- Upgrade V2 -> V3: V3 doesn't know about the transcripts anymore (the
field is removed). V2 created a summary with `None` transcripts meaning
their hash was not included, good.
- Rollback V3 -> V2: V3 unconditionally sets the sentinel to `true`,
meaning V2 will interpret the transcripts as `None` and not try to hash
them.
 - V3 <-> V4 ([dfinity#11224](dfinity#11224)):
     - The goal of V4 is to remove the introduced sentinel field.
     - Upgrade V3 -> V4: V4 will ignore the sentinel field set by V3.
- Rollback V4 -> V3: V3 will deserialize a summary with an unset
sentinel, but it does not read it anyways.