[testnet] ScyllaDB: eliminate per-batch read in write_batch via two-phase writes (take 2)

[ma2bd]

Motivation

Reland of #6334 (merged then reverted), now fixing the atomicity regression that the first attempt introduced.

Previously, every ScyllaDB write_batch containing a DeletePrefix that overlapped a Put in the same batch paid a find_keys_by_prefix read inside UnorderedBatch::expand_colliding_prefix_deletions. The expansion existed because all statements in a CQL unlogged batch share one write timestamp by default, and at equal timestamps a range tombstone shadows a same-batch insert. This fired on the hot path (notably reset_chain_manager clearing pending_*_blobs views).

Proposal

Resolve the in-batch prefix/insert collision with explicit write timestamps instead of a read, while preserving write_batch atomicity.

• Exclusive mode (single writer per partition): we own the write timestamps, so we issue the whole write as a single atomic unlogged batch with explicit per-statement USING TIMESTAMP — prefix-deletions at T, single-key deletions, insertions, and a sentinel at T+1. The higher data timestamp keeps a range tombstone from shadowing a same-batch insert, and the intended ordering is pinned by the timestamps rather than by send order. The per-store timestamp floor is seeded once on first write by reading the WRITETIME of a reserved sentinel row (empty clustering key, unused by views and journaling alike), then advanced monotonically.
• Shared mode (coordinator-assigned timestamps): we split the write into two sequential unlogged batches — phase 1 the prefix-deletes, phase 2 the puts and single-key deletes.

This supersedes the reverted #6334: that version also split the exclusive write into two sequential batches, which broke atomicity — a failure between phases could leave prefixes deleted but their replacement data missing. Since exclusive mode controls the timestamps, the ordering no longer depends on issuing two batches, so the whole write is now one atomic batch.

Drop the now-unused expand_colliding_prefix_deletions. UnorderedBatch's SimplifiedBatch::from_batch becomes pure (Ok(batch.simplify())), so the per-batch read disappears at the journaling-layer boundary. SimpleUnorderedBatch is unchanged and still used by DynamoDB, which needs full expansion as it has no native range-delete primitive.

Test Plan

CI, plus new exclusive-mode tests in linera-views/tests/store_tests.rs:

• test_scylla_db_writes_from_state_exclusive — routes the DeletePrefix-overlapping-Put cases through the explicit T/T+1 path.
• test_scylla_db_exclusive_seed_after_restart — checks that a reconnected store reseeds its timestamp floor from the persisted sentinel and resumes strictly above prior data.

Release Plan

front-port to main