[opt](point query) reduce short circuit lookup cache usage (#67228) - #69020

)

LookupConnectionCache preallocates 32 reusable Blocks for each cached
point-query context. For wide tables, these Blocks dominate the cache
memory usage, and the overhead grows significantly with the number of
client connections.
In the current implementation, a Reusable instance is not shared across
client connections, and requests within the same connection are executed
serially. Therefore, the concurrency of each Reusable instance does not
exceed one, making it unnecessary to preallocate 32 Blocks.
This change makes the Block pool size configurable and reduces the
default size from 32 to 1. If concurrent access is introduced in the
future, additional Blocks can still be allocated on demand, while the
pool retains only the configured number of idle Blocks.
In a real production online point-query workload querying a wide table
with about 80 columns through approximately 8,000 client connections,
the LookupConnectionCache memory usage on a single BE was reduced from
about 10 GB to 1.8 GB.

Co-authored-by: TRAE CLI <[email protected]>