Fill the gaps in RollingSessionWindow from the runtime state

[ordian]

This is a follow-up to #6040 aiming to address an issue described in #6106 (comment).

The edge case scenario can arise when 1/3+ of active validators are being offline or somehow lose their parachain db for a few sessions (> dispute_period, currently 6). We don't expect this to happen on mainnet, but it could be an annoyance for small testnets with short session durations.

With #6106, validators will try to persist the session window, but if there are missing sessions at the beginning of the window, when querying runtime, they will use the state of the best block typically, which only contains the latest dispute_period sessions.

A way to address this would be to modify the API to allow the caller to pass the appropriate (old) relay parent needed to query the old SessionInfos.

[ordian]

There cache_session_window_for_head API already exists, but currently, it's only being called for active leaves in approval-voting and dispute-coordinator.

[rphmeier]

A way to address this would be to modify the API to allow the caller to pass the appropriate (old) relay parent needed to query the old SessionInfos.

Old relay-parents can't be relied upon to have state available. If anything, a few hundred blocks behind the finalized block is the limit. How old are you thinking?

[ordian]

a few hundred blocks behind the finalized block

We are only concerned about unfinalized blocks here. In case finality stalls for a few sessions for some reason.

[sandreim]

This was implemented when removing RollingSessionWindow, see #7123 and #6968