slow query log: only use Instant::now when needed + refactor

[Centril]

Description of Changes

• Refactor the slow query logger to reuse WorkloadType and other refactoring.
• Try to regain some of the performance lost in incr-select (see below).

After 7d5eb15 compared to 7d5eb1532db173c032afa6b46b013148aec132c5~1:

Benchmarking incr-select: Collecting 100 samples in estimated 5.0
incr-select             time:   [226.69 ns 228.30 ns 230.24 ns]
                        change: [+24.989% +26.097% +27.045%] (p = 0.00 < 0.05)
                        Performance has regressed.

After this PR compared to 7d5eb1532db173c032afa6b46b013148aec132c5~1:

Benchmarking incr-select: Collecting 100 samples in estimated 5.0
incr-select             time:   [195.23 ns 195.55 ns 196.04 ns]
                        change: [+7.5233% +8.3927% +9.1255%] (p = 0.00 < 0.05)
                        Performance has regressed.

Run on i7-7770K, 64GB RAM.
As you can see, this recovers some of the +26% lost in performance such that +8.4% is lost instead.
The primary reason for the performance loss was the unconditional Instant::now which results in a system call.
This call is now conditional on whether there's also a threshold.

API and ABI breaking changes

None

Expected complexity level and risk

1

Testing

No semantic changes.

[mamcx]

LGTM

[joshua-spacetime]

Thanks for catching this @Centril. I'll take a look shortly. You're correct that we shouldn't be making any system calls per eval_incr.

[joshua-spacetime]

Thanks @Centril. I'm not sure where that 8% regression is coming from. It doesn't appear we've added any significant work.

I did notice though that we only ever call SlowQueryLogger::query, but we should be calling SlowQueryLogger::subscription in eval and SlowQueryLogger::incremental_updates in eval_incr. Could you make those changes here? It should only be two lines.

[Centril]

So I commented out slow_query.log(); and replaced it with black_box(slow_query). This reduces the regression to about 5%. Further commenting out let slow_query = SlowQueryLogger::incremental_updates(ctx, sql); makes the regression go away.