Add idle session cleanup to prevent unbounded session accumulation

[ddworken]

Summary

• Added automatic cleanup of idle sessions when session count exceeds threshold
• Sessions inactive longer than idle timeout are terminated to free resources
• Prevents memory exhaustion from clients that create sessions but never terminate them

Changes

• Added configurable session idle timeout (default 30 minutes)
• Added background cleanup task that activates when session count > 10k
• Track last activity time for each session
• Double-check idle status before cleanup to prevent race conditions

Testing

• Added tests for idle session cleanup behavior
• Added tests for threshold-based activation
• Added tests for session activity tracking
• All existing tests continue to pass

This ensures the server remains stable even when clients don't properly close their sessions.

[ochafik]

Thanks @ddworken for preparing this!

I'd feel more comfortable if we limited the number of open sessions first (and decline further sessions w/ a 503). There's no good way to label long sessions as inactive atm, as there are growing use cases (and spec change proposals) around long running tool calls for instance, so 30 min w/o activity may be perfectly fine / ditching these sessions would be a breaking change.

Also not sure the current changes prevent a crash caused by a very high number of sessions started at once?

[ochafik]

long running tool calls for instance, so 30 min w/o activity may be perfectly fine / ditching these sessions would be a breaking change.

Other suggestions:

• Define idle as "no activity since N seconds and no pending request": while this would allow long running tools, it would still not allow patterns of, say, a client that's just listening to resource updates all day long (and reads said updated resources once a day to index them), so I probably wouldn't default to this either
• Define idle as "no open connection / stream since N seconds" (i.e. connection was closed and not resumed for a while). Would keep track of # open streams for a given connection (maybe in the try/finally of run_server). Probably the least breaking change

WDYT?

[felixweinberger]

Related: #1159