The main branch is slower than expected

[wenhuach21]

Quantizing llama3-8B cost 15+, expected is 13

[xin3he]

Root cause: #944

[xin3he]

auto-round --model /models/Meta-Llama-3.1-8B-Instruct/ --scheme "W4A16" --device_map 1 --enable_torch_compile

[wenhuach21]

1 The torch.compile change should also be reverted, as it introduces approximately 2 minutes of tuning overhead (@xin3he, see #987).

2 The 1.5-minute overhead introduced in #944 needs to be root-caused.

3 The immediate_packing implementation introduced a 1-minute overhead, which may be mitigated by postponing the CPU transfer until the packing phase.

The target total tuning time is 13 minutes.