[Draft][triton] Implemented 8bit optimizers on XPU

[Egor-Krivov]

Implemented 8bit optimizers in triton to use of XPU devices.

Currently there is no interface for kernel registration of 8bit optimizers, so I just hardcoded usage for testing.

Tested with BNB_TEST_DEVICE="xpu" pytest --show-capture=no -q tests/test_optim.py::test_optimizer8bit

Benchmarked essentially on the same test, getting better performance than torch optimizer.

This PR contains 3 implementations:

1. Pure torch implementation that materializes during quantization, hence large memory usage. Can be used for testing purposes.
2. Combination of torch and triton kernels for quantization+dequantization
3. Pure triton implementation - fastest, used currenlty.