refactor!: avoid hard dep to tokio rt

[tisonkun]

This closes #3948.

I'm trying to opt-out tokio deps as much as possible. But it surprises me that simply remove "rt" doesn't give compile error locally. Check CI workflow for more infos.

[Xuanwo]

We need to make blocking layer under a feature gate instead. This is a breaking change.

[tisonkun]

@Xuanwo I have an alternative idea.

I'm investigating pollster::block_on to replace tokio runtime. It's a trivial "runtime" to block on the current thread exactly. But I noticed that we have a few features that depends on a running tokio runtime (e.g., pg services), so we may still depends on tokio runtime for running a compatible blocking layer.

[Xuanwo]

I'm investigating pollster::block_on to replace tokio runtime.

Replacing tokio this way doesn't look good to me. I like the idea of using pollster, but I wish it to be a new layer like PollsterLayer instead.

In this way, our users can decide which runtime to use by switching layers in the next version.

[tisonkun]

so we may still depends on tokio runtime for running a compatible blocking layer.

This is the case. Switching to a feature flag solution.

[tisonkun]

@Xuanwo @PsiACE seems ready for review now.

[Xuanwo]

Mostly LGTM, the only thing here is to add a upgrade entry in UPGRADE.md since this is a breaking change.

[Xuanwo]

Thanks!