PrioGraphScheduler::try_schedule_transaction remove TransactionAccountLocks allocation

[apfitzge]

Problem

• get_account_locks_unchecked does some allocations for locks
• try_lock_accounts takes arbitrary iterators (clone-able) so we do not need to allocate

Summary of Changes

• Make AccountKeys::key_segment_iter return clonable iterator. (No change to impl, just exposes clone)
• Make AccountKeys::iter return clonable iterator. (No change to impl, just exposes clone)
• Iterate over write/read accounts, do not allocate.

Fixes #

[apfitzge]

Small part of larger effort to reduce allocations. Don't expect significant impact from this alone.

Banking Bench

master:
{'name': 'banking_bench_total', 'median': '51532.09'}
{'name': 'banking_bench_tx_total', 'median': '53217.47'}
{'name': 'banking_bench_success_tx_total', 'median': '50988.14'}

with change:
{'name': 'banking_bench_total', 'median': '51722.87'}
{'name': 'banking_bench_tx_total', 'median': '53396.77'}
{'name': 'banking_bench_success_tx_total', 'median': '51339.26'}

bench-tps

master:
[2024-06-17T14:04:13.835628171Z INFO  solana_bench_tps::bench] 
    Average max TPS: 58954.28, 0 nodes had 0 TPS
[2024-06-17T14:04:13.835631979Z INFO  solana_bench_tps::bench] 
    Highest TPS: 58954.28 sampling period 1s max transactions: 3346725 clients: 1 drop rate: 0.58
[2024-06-17T14:04:13.835635625Z INFO  solana_bench_tps::bench] 	Average TPS: 36731.09

with change:
[2024-06-17T13:59:15.938553191Z INFO  solana_bench_tps::bench] 
    Average max TPS: 61463.46, 0 nodes had 0 TPS
[2024-06-17T13:59:15.938564112Z INFO  solana_bench_tps::bench] 
    Highest TPS: 61463.46 sampling period 1s max transactions: 3123034 clients: 1 drop rate: 0.62
[2024-06-17T13:59:15.938578429Z INFO  solana_bench_tps::bench] 	Average TPS: 34279.95

left is with change, right without.

[apfitzge]

@t-nelson - want to make sure I'm not an idiot with the SDK changes.

As far as I know, this is not a breaking change since I am only adding an additional trait-bound on the return type. It's not actually changing the return type, just exposing to calling-code that it is also Clone-able.

[ryoqun]

As far as I know, this is not a breaking change since I am only adding an additional trait-bound on the return type. It's not actually changing the return type, just exposing to calling-code that it is also Clone-able.

(yeah, i think this is correct)

[ryoqun]

(yeah, i think this is correct)

oh, one little caveat is that we're now committed to returning cloneable iter thereafter. but, i don't think this will be a significant burden.

[ryoqun]

lgtm