RESP3 when using pub/sub can cause unexpected connection closures due to backlog #3263

Whenever a connection smells of pub/sub, Redis applies different output buffer limits:

This means that any of:

• large BLOB fetches (over the 32MiB limit)
• bursts of pipelined traffic (over the 32MiB limit)
• sustained regular traffic (over the 8MiB limit for 60s)

will trigger spontaneous connection death, even when the cause has nothing whatsoever to do with pub/sub.

This has been a latent risk since we enabled RESP3 by default, but was masked because until #3255 we had a glitch whereby RESP3 connections didn't subscribe the legacy config-channel; this fix in main has not yet been released to NuGet, but releasing this would expose multiple consumers to risk.

I DO NOT PLAN TO RELEASE #3255 UNTIL THIS IS RESOLVED

Current plan

• Introduce a new config option that enables shared pub/sub interactive connections on RESP3, that defaults to false
• Update the docs and tests to not claim that we expect RESP3 to reduce connections unless this flag is enabled (also noting that you might want to check your server output limits in that case)