torch compile and cuda alloc improvements

[winglian]

add back missing torch.compile to hf trainer set cuda allocation env var to improve llama vram use by up to 17%