[Unity] Add Relax multi-device e2e cases

[yongwww]

This Pull Request enables Relax multi-device end-to-end test cases, including multi-CPU, multi-GPU, and heterogeneous CPU-GPU cases.

• Add the kTarget attribute to prim_func during LegalizeOps if the related vdevice is annotated.
• Change CallTIRRewrite pass to ModulePass, find and fill the device_index in the emitted alloc_tensor by looking up the global vdevice list of context mod.
• Update DefaultGPUSchedule to only works for prim_funcs that are supposed to be on GPU, allowing prim_funcs to work for multiple targets.
• Make the target argument of relax.build optional, and use the annotated target in the vdevice list for compilation if it is not specified.

cc: @tqchen @psrivas2 @junrushao @csullivan @Lunderberg @zxybazh @qelk123

[tqchen]

cc @zxybazh @Hzfengsy

[zxybazh]

LGTM, thanks @yongwww!