fix: accelerate the first inference speed on low-level NPUs.

[muxuezzz]

This PR will implement a binary mode to accelerate the first inference speed on NPUs.
When using the NPU:

• For Atlas A2 training series products, jit_compile is disabled by default (jit_compile=False).
• For Atlas training series products/Atlas inference series products, jit_compile is enabled by default (jit_compile=True).

This causes chips like the 310P to compile required operators in real-time during the first inference rather than using existing operators from the operator library. As a result, ComfyUI becomes extremely slow during the first generation. For example, with SD1.5, the first generation takes approximately 600 seconds without setting jit_compile=False, but only about 40 seconds after setting it to False.

[Kosinkadink]

Are there any performance penalties if jit_compile is set to False for the second run onwards, or is the sampling of subsequent runs the same regardless of jit_compile setting?

If there are not performance downsides to flipping it off, we'll merge this in, but I'd like your confirmation first!