🐛 Bug

Enabling autograd.profiler introduces nearly a 180% overhead (4.6s v. 8.17s)

To Reproduce

1. From benchmarks dir, run numactl -C 3 python -m fastrnns.bench --fuser=te --executor=profiling --group=rnns --rnns=jit_premul_bias to get a baseline
2. Apply the following patch to instrument the bench with autograd.profiler: https://github.com/pytorch/pytorch/commit/dd4b3268605c055a1a8653f8554ccffc9e874044.patch (or git cherry-pick dd4b3268605c055a1a8653f8554ccffc9e874044
3. Re-run the bench (no re-comp is necessary as the patch contains only python changes).

Expected behavior

An overhead that's within 5-10%.

@ngimel @ilia-cher

cc @VitalyFedyunin @ngimel