Add xpu version for async offload and some other things.

[simonlui]

Besides adding async offload for XPU, I also did the following:

• Cleaned up some xpu memory management and gated it by Pytorch version since it isn't clear what the minimum version of Pytorch ComfyUI actually enforces but I did see 2.3 used at the time of writing this so I went with that as kind of the minimum baseline for what needed a check and didn't.
• Changed the checks for BF16 and FP16 on XPU to be more precise using device properties that exist now which can be queried and again, gated by Pytorch version.
• Turned on non-blocking for XPU, this no longer freezes since around a year ago and speeds up stuff.

Again, works for my machine but ideally would get tested by others.