Fix LARS/LAMB optimizer support and non-contiguous tensor handling on XPU

[jiqing-feng]

Changes

• Enable LARS/LAMB optimizers on XPU: Added missing "lars" and "lamb" entries to name2optimizer_id, name2optimizer_32bit_fn, and name2optimizer_fn dicts in the Triton backend, and name2optimizer_id in the default backend. LARS maps to MOMENTUM (1-state) and LAMB maps to ADAM (2-state).

• Fix Triton compilation error with fp16 gradients: In _optimizer_precondition_1state_32bit, the MOMENTUM branch's step == 1 path does s1_vals = g_vals (direct assignment). When gradients are fp16, this changes s1_vals from fp32 to fp16, conflicting with the else branch where arithmetic auto-promotes to fp32. Fixed by casting g_vals to fp32 at the assignment site.

• Fix non-contiguous tensor support for blockwise quantization: Triton kernels use linear offsets to access memory, which is incorrect for non-contiguous tensors. Added .contiguous() calls in quantize_blockwise and dequantize_blockwise entry points.

Related tests:

pytest -k "xpu" -ra test_ops.py::TestNonContiguousInputs::test_quantize_blockwise_non_contiguous
pytest -k "xpu" -ra test_optim.py::test_optimizer32bit

Hi @matthewdouglas . Would you please review this PR? Thanks!