[CAE-1046] fix(loading): cache the loaded flag for slave nodes

[ndyakov]

Checking the loading state by sending a ping on each node selections increases the latency. See comments in #3370 .
This introduces a naive cache that should improve the latency once the node is marked as loaded

[Copilot]

Pull Request Overview

Adds a caching mechanism to skip repeated loading checks for slave nodes once they’re determined to be “loaded.”

• Introduces an atomic loaded flag to clusterNode
• Resets loaded when a node is marked as failing
• Updates Loading() to short-circuit and set loaded after the first non-loading result

Comments suppressed due to low confidence (1)

osscluster.go:343

• [nitpick] The field name loaded is a bit ambiguous. Consider renaming it to something like isLoadComplete or loadCached to clarify that it’s a one-time cache flag.

loaded     uint32 // atomic

[LINKIWI]

Thanks! I agree this should resolve the latency regression introduced in v9.9.0 for the typical case slave request.

Ideally (in the future) we can update the loading state as a side effect of user command execution, in order to obviate the need for an explicit additional critical-path PING.