go1.21: use context.WithoutCancel to avoid propagating user-initiated cancelations

[aarongable]

As of #5459, requests which arrive at the SA have their ctx cleared and replaced after the SA starts processing the request. This is to work around an issue where context cancellation (e.g. due to a deadline/timeout coming up, or due to the initiating gRPC request being closed) propagating into sql query cancellation causes sql connections to be marked as bad and dropped from the pool, resulting in connection churn.

However, this prevents the context object from carrying trace information as it moves through the SA; the SA cannot add custom trace data or additional spans, because the context object has been cleared of that data.

We should investigate other ways to prevent cancellation of sql queries that would allow the incoming context object to be propagated throughout the SA to a greater degree.

[aarongable]

When go1.21 or go1.22 is released, we should use the upcoming context.Detach API to accomplish this.

[jsha]

is to work around an issue where context cancellation (e.g. due to a deadline/timeout coming up, or due to the initiating gRPC request being closed) propagating into sql query cancellation causes sql connections to be marked as bad and dropped from the pool, resulting in connection churn.

Note that this impact on connection churn is still hypothesized - we haven't yet been able to prove that this is a cause of connection churn.

Also, an update: we now "detach" contexts at the WFE instead of the SA.

[aarongable]

go1.21 is shipping with context.WithoutCancel! https://tip.golang.org/doc/go1.21

We may also be able to handle some of our use-cases by using the new context.AfterFunc to register cleanup functions that should be run after the context is canceled.