Add LARS to str2optimizer32bit dictionary

LARS optimizer was missing from str2optimizer32bit, causing KeyError
when using LARS32bit optimizer. LARS uses momentum-based kernels since
it's essentially SGD with momentum plus layerwise adaptive learning rates.

Fixes #1810

Author: Mr-Neutr0n

bitsandbytes/backends/cuda/ops.py

@@ -574,6 +574,10 @@ def _gemv_4bit_impl(
         lib.cademamix32bit_grad_fp16,
         lib.cademamix32bit_grad_bf16,
     ),
+    "lars": (
+        lib.cmomentum32bit_grad_32,
+        lib.cmomentum32bit_grad_16,
+    ),
 }
 
 str2optimizer8bit_blockwise = {