Change CanonicalState sharing model.

[deuszx]

Motivation

The exporter's CanonicalState is cloned via clone_unchecked(), giving the exporter a LogView with a frozen stored_count. After flush persists entries to DB, the exporter's LogView still can't read them – LogView::get() routes index >= stored_count to new_values (empty in the clone), never hitting the DB. This means the in-memory buffer can never be cleared and grows without bound.

Proposal

Share a single CanonicalState behind Arc<tokio::sync::RwLock> instead of cloning it. Both the block processor and exporter use the same LogView, so stored_count stays consistent. After flush +post_save(), the buffer is safely drained — subsequent reads fall through to the LogView which has the correct stored_count.

CanonicalState

• Removed flushed_count field (flush now drains the entire buffer)
• Replaced Arc<papaya::HashMap<usize, CanonicalBlock>> with Vec<CanonicalBlock> (RwLock provides concurrency; entries are sequential)
• Removed clone() (no more clone_unchecked()) and next_index()
• flush() now uses buffer.drain(..) instead of iterating flushed_count..count

SharedStorage

• shared_canonical_state field is now Arc<tokio::sync::RwLock<CanonicalState<C>>>
• clone() is infallible (Arc clone) — kept Result return type to avoid touching 13+ call sites
• push_block() is now async (acquires write lock)

ExporterStorage / BlockProcessorStorage

• Read operations (get_block_with_blob_ids, get_block_with_blobs, get_latest_index) acquire a read lock
• Write operations (push_block, save) acquire a write lock
• clone() takes &self instead of &mut self

Test Plan

• cargo build -p linera-service — compiles with no errors
• cargo test -p linera-service --bin linera-exporter — all existing tests pass (including test_committee_destination)

Release Plan

• These changes should be released to testnet, and then
• Backported to main.

Links

• reviewer checklist

[deuszx]

Superseeded by #5449