Update the benchmarking script to accept an arbitrary BaseImageAugmentationLayer subclass

[LukeWood]

This will allow us to make the decision on whether or not to JIT compile specific layers

[bhack]

Also when will we have the new benchmark I hope that:

• We still let the user to control XLA jit compilation as discussed in jit_compile control in BaseImageAugmentationLayer keras-io#1541
• We could maintain an updated inventory .txt or .MD in the repository for the layers that could not be compiled for missing XLA ops coverage as in the last point at vectorize KPLS: tf.vectorized_map fallback cause #291 (comment)

[LukeWood]

Closing this for now, we do some benchmarking internally, but I don't think we want to benchmark these at this time due to resource constraints.