remove use of nanobind for _export pass with shardy
right now using these nanobind methods requires going to bytecode and back. This creates extra overhead for these simple operations. Moving these to pure python removes extra lag.
lowered_with_shardy and get_mesh are 1 to 1 reproduced in Python.
sdy_round_trip_import_shardings is used from xla_sdy_capi
with change gain 18% improvement on export of jitted function.
PiperOrigin-RevId: 775439256
