Using object.size() in init.R could be expensive

Describe the bug

87f9bb6f8b450d897fb191780b086de07805fd4e in #544 uses object.size(obj) in inspect_env in init.R to support showing object size in the workspace viewer. However, the performance of this function could be extremely poor if the object contains character vectors in it.

library(data.table)
dt <- data.table(id = 1:5000000)
for (i in 1:20) {
  dt[, paste0("x", i) := rep("hello", .N)]
}
system.time(object.size(dt))

   user  system elapsed
  1.068   0.687   1.741

This means, with session watcher enabled, whenever this data table is present in the global environment, then user has to wait for 1.7s after evaluating each top-level expression.

Therefore, I don’t think we should call object.size() so eagerly in this way.

[renkun-ken]

With #581 merged, now the object.size() is called less frequently: it is called only when a new symbol appears or its memory address or length has changed.

And we have an option vsc.show_object_size to opt-in. It is disabled by default to minimize possible delay.

[renkun-ken]

Or we could estimate it by getting the size of a subset of a given object?

It might add too much complexity if we take into account nested list, object attributes, etc. where character vector could appear anywhere. If we don’t handle these cases, it won’t work, e.g. a multi-level nested list with some big character vectors will trigger long waits.

Also, if we had a robust way to do this, the object size might not be useful as it omits the size of character vectors, making the result misleading.