improve replay performance

[smcio]

Quite a few changes here to improve perf. These changes improve perf by quite a bit, including cutting block exec time down by about 50% in many cases.

A list of changes:

• Cache each transaction’s message identity so it only needs to be serialised and hashed once during replay.
• Speed up large AccountsDB reads with batched index lookups and parallel iterators.
• Add a bounded decoded-account cache which only retains accounts that are actually being reused.
• Make the decoded-account cache size configurable with tuning.common_account_cache_max_mb.
• Split account overlays into shards so parallel transactions no longer contend on one global lock.
• Publish touched transaction accounts in batches, with the same tombstone and last-write-wins behaviour as before.
• Reduce replay setup work by pre-sizing maps, avoiding repeated account lookups and doing less duplicate bookkeeping.
• Rework LtHash calculation to reuse buffers and spread larger workloads across workers.
• Cache validator sets and certificate verifiers used for Alpenglow vote rewards.
• Keep epoch stake snapshots immutable and retain the exact vote account selected by the leader schedule.
• Add more detailed replay timings for account loading, transaction publication, rewards and bank hashing.
• Expand the replay timing viewer so the new measurements are easier to break down.
• Fix epoch-boundary inflation handling when the first slots of an epoch are skipped.
• Remove the unsafe overlay bank-hash shortcut and keep the canonical modified-account journal as the source of truth.
• Add regression, concurrency, race and benchmark coverage for the new replay paths.