Restructure cleaning up of the futures in decoupled mode

[Tabrizian]

Fixes triton-inference-server/server#6270 (comment)

Huge shout out to @rmccorm4 for finding the root cause.

[rmccorm4]

Nice fix! Spoke offline, let's try to make the messages in these blocks also unique_ptr instead of shared_ptr if possible by moving the unique_ptr into the lambda captures, to reduce risk of reintroducing a leak due to long lived reference.

[Tabrizian]

@rmccorm4 I tried using unique_ptr but for some reason it looks like it needs to be a const unique_ptr so it looks like it might need further digging to make it work with unique_ptr.

[rmccorm4]

@rmccorm4 I tried using unique_ptr but for some reason it looks like it needs to be a const unique_ptr so it looks like it might need further digging to make it work with unique_ptr.

Can you add a ticket to look into it later? I think it would be good to reduce future risk.

[krishung5]

Nice work on finding the root cause and fixing the leak!

[Tabrizian]

@rmccorm4 Filed a ticket.