perf: get account and access key in parallel

[mooori]

At high transaction throughput, get_account and get_access_key are two of the most expensive items in Runtime::Apply. See for example this profile of apply-range from January 2025.

These calls to get_account, get_access_key happen sequentially in verify_and_charge_transaction. Getting them in parallel should shave some time of the critical path. With TrieUpdate becoming Send + Sync [#12981], this might become achievable.

[mooori]

Updates:

• Simple approach using rayon::join degrades performance.
• A more sophisticated approach could include worker threads that send/receive work over channels. At the moment TrieUpdate is Send + Sync only for immutable references, which makes this approach impractical.
• Another approach could be based on getting data for more than one transaction in parallel, see below. Note that this conflicts with other work (e.g. #12983).

take the first n txs such that they all have different signers
prefetch (account, access_key) for these n accounts in parallel (ok as all signers are different)
process the transactions, passing in the prefetched data
repeat until all transactions are processed

This probably degrades performance when a chunk contains many txs coming from the same signer, which will make all n small. As a mitigation, if n is mostly small for a list of transactions then this prefetching could be disabled. Should be a O(num_txs) pre-check.