execution: fix OOM crash due to db::Buffer::accounts realloc (#2064)

[battlmonstr]

The flat_hash_map has a policy (in rehash_and_grow_if_necessary()) to grow 2x when its size reaches 25/32 (78%) of capacity.

Previously the batch execution logic was adding things to db::Buffer, and then testing if its size becomes bigger than batch_size. This lead to significant overshooting, because at the growth threshold it will double the memory size and make it batch_size * 2 (even more, because there's also unused 22% of capacity).

In standalone silkworm execution stage the batch_size estimation was based on a gas-based heuristic formula, which lead to significant underestimation and an OOM crash when trying to allocate 4 Gb of RAM at once.

This changes the logic from checking batch_size aposteriori to apriori. If there's a possibility to grow the buffer significantly and overshoot the limit, we shouldn't take the risk, and commit the batch.

Additional changes:

• execution stage: flush history after each block as in CAPI
• capi: persist the executed work before quitting on invalid block
• remove unused db::Buffer::current_batch_history_size()
• ExecutionProcessor: split flush_state from execute_block
• block_buffer.terminate_and_release_all using gsl::finally

[canepat]

We should change this to something like:

[[nodiscard]] FlushingResult flush_state() noexcept;

in order to:

1. make ExecutionProcessor exception-free again
2. do not rely on db::Buffer::MemoryLimitError exception to handle an expected condition i.e. batch full

[canepat]

Closes #2064