ENH: Benchmark different backends and compare them to pure implementations

[Dekermanjian]

Context for the issue:

Add an example of how Pytensor backend switching compares to a pure implementation in that backend.