[torch_xla2] Use torch.Generator for random number generation in Environment

[will-cromar]

• Pull jax.random.keys from torch.Generator (either argument or global generator)
• Use randn as a test case
  • Fix empty because I thought it was required for randn (it's probably not)
• Support nondeterministic output in test_ops
  • Bonus: use torch.testing.assert_close because it prints a much better error message upon failure; instead of False is not True, you get:

E       AssertionError: Scalars are not close!
E       
E       Expected 0 but got 140357434826080.
E       Absolute difference: 140357434826080 (up to 0.001 allowed)
E       Relative difference: inf (up to 1e-05 allowed)

• Remove manually-managed PRNG key