[Bug]: AutoDeploy: avoid aot compilation of auto_deploy kernels unless kernel is needed

[lucaslie]

System Info

see https://nvidia.slack.com/archives/C059LSY62BT/p1762167016530579

Right now, our moe_align extension gets compiled when auto_deploy gets imported. This can affect tooling that imports auto_deploy even if auto_deploy is not used or when the moe_align extension is not used.

Let's minimize the times that the extension gets imported and hence compiled. A couple of suggestion for that:

• Use local imports inside functions instead of top-level file imports
• Avoid auto-importing the moe_align or any other extension