Performance regression with --gc:markandsweep

[rbehrends]

As a result of this commit, the GC can be triggered for every allocation in a series of small allocations if compiled with --gc:markandsweep, leading to massive overhead.

Example

For example, on my machine the following code runs in 50ms with the default GC, but takes nearly 4 seconds with --gc:markandsweep.

const n = 1_000_000

type T =
  ref object
    val: int
    next: T

proc main =
  var t: T = nil
  for i in 1..n:
    t = T(val: i, next: t)

main()

[timotheecour]

maybe related to #10040