Disable host callbacks and enforce 32bit precision when running on TPU

[flferretti]

This PR introduces two main changes to allow runnning simulations on TPUs:

1. host callbacks are disable by default, since they are not fully supported for TPU architectures
2. 32bit precision is enforced when importing jaxsim

Example error on js.model.step:

[/usr/local/lib/python3.10/dist-packages/jax/_src/interpreters/pxla.py](https://localhost:8080/#) in __call__(self, *args)
   1275           or self.has_host_callbacks):
   1276         input_bufs = self._add_tokens_to_inputs(input_bufs)
-> 1277         results = self.xla_executable.execute_sharded(
   1278             input_bufs, with_tokens=True
   1279         )

XlaRuntimeError: INVALID_ARGUMENT: Unknown send channel 33 for host callbacks

---

📚 Documentation preview 📚: https://jaxsim--263.org.readthedocs.build//263/

[diegoferigo]

Thanks @flferretti! Let me know what do you think of the suggestions below, there were cases that could be handled differently from what you proposed.

Please wait #262 to be merged before this one.