Improve `write_shared` performance

[GuillaumeGomez]

As explained in this comment, we could improve performance one step further by better use serde_json API.

[yotamofek]

Just a random semi-unrelated note:

something that might really improve perf would be the ability to serialize an impl Display into a JSON string. serde doesn't support that, we need to first render out all fields into Strings and only then serialize them, which is probably very wasteful.

I don't know if there's an alternative JSON serializer that does support such a use case, but if anyone has an idea - I think it would be a really cool experiment. We have quite a few places, IIRC, where that could obviate large allocation (which would then be tacked on to other large allocations, causing more reallocations, etc...)

[yotamofek]

Just found out about Serializer::collect_str, looks promising!

[GuillaumeGomez]

Sounds interesting indeed!