fix #16752: threadvar now works with importcpp types; osx now uses native TLS (--tlsEmulation:off), which can be orders of magnitude faster (#16750)

* osx now uses native TLS, which can be orders of magnitude faster

* add {.cppNonPod.}

* improve test

* changelog, docs, disable part of windows test

tests/benchmarks/readme.md

@@ -0,0 +1,5 @@
+# Collection of benchmarks
+
+In future work, benchmarks can be added to CI, but for now we provide benchmarks that can be run locally.
+
+See RFC: https://github.com/timotheecour/Nim/issues/425

tests/benchmarks/ttls.nim

@@ -0,0 +1,30 @@
+discard """
+  action: compile
+"""
+
+#[
+## on osx
+nim r -d:danger --threads --tlsEmulation:off tests/benchmarks/ttls.nim
+9.999999999992654e-07
+
+ditto with `--tlsEmulation:on`:
+0.216999
+]#
+
+import times
+
+proc main2(): int =
+  var g0 {.threadvar.}: int
+  g0.inc
+  result = g0
+
+proc main =
+  let n = 100_000_000
+  var c = 0
+  let t = cpuTime()
+  for i in 0..<n:
+    c += main2()
+  let t2 = cpuTime() - t
+  doAssert c != 0
+  echo t2
+main()