[Accelerate] allow get_execution_device to be used when initializing a model

[kylesayrs]

Purpose

• Extend get_execution_device to be usable when initializing a model in the init_empty_weights context
  • This is used by [Performance] Add memory compression and decompression pathways #301 when initializing the model structure with CompressedLinear layers

Changes

• Remove meta device check, which fails if in the init_empty_weights context
• Make execution device logic more explicit, implement fallback for modules which do not have parameters

Testing

• Added tests to check behavior under the init_empty_weights context