parameterized test_graph_optims and test_graph_scaling_fused_optimizers (#133749)

Fixes #123451

This is a rework of a reverted pull request, #125127.
The test failure is fixed.

Pull Request resolved: #133749
Approved by: https://github.com/janeyx99

Author: zero000064

test/inductor/test_compiled_optimizers.py

@@ -146,6 +146,8 @@ class KernelCounts(NamedTuple):
     "test_sgd_weight_decay_maximize_cuda": 4,
     "test_sgd_weight_decay_maximize_xpu": 4,
     "test_sgd_weight_decay_maximize_cpu": 4,
+    "test_sgd_weight_decay_cpu": 4,
+    "test_sgd_weight_decay_cuda": 4,
     "test_sgd_momentum_weight_decay_foreach_cuda": 2,
     "test_sgd_momentum_weight_decay_foreach_xpu": 2,
     "test_sgd_momentum_nesterov_weight_decay_foreach_cuda": 2,