Box NewTaskPayload to reduce size greatly

[ryoqun]

Problem

NewTaskPayload is large (more than 200 bytes) due to its direct inclusion of ExecuteTimings. In turn, this causes large mem writes/reads around crossbeam channel even for NewTaskPayload::Task.

Note that this is particularly problematic for block production mode (senders are multi-threaded whereas block verification is currently single-threaded) because those writes are inside crossbeam channel's linearization point, creating a source of contention.

Summary of Changes

Box the rare variant, which only occurs once per bank, benefiting the other variant channel transmission (99.9% of time). Overall, extra indirection for access and mem alloc/dealloc for the NewTaskPayload::OpenSubchannel variant pays off.

before

ledger processed in 13 seconds, 977 ms
ledger processed in 13 seconds, 777 ms
ledger processed in 13 seconds, 905 ms

after

maybe 1-2% gain.

ledger processed in 13 seconds, 678 ms
ledger processed in 13 seconds, 512 ms
ledger processed in 13 seconds, 856 ms

Extracted from: #2325