Standardize weight_decay_mask naming and add weight decay

[rdyro]

Test PR co-written with Gemini.

Addresses inconsistencies where the mask parameter for weight decay was named mask instead of weight_decay_mask in optimizers like adamw, nadamw, adan, lion, lamb, and adamaxw. Additionally, missing decoupled weight decay support (with weight_decay and weight_decay_mask) was added to multiple base optimizers including adabelief, adagrad, amsgrad, radam, rmsprop, sgd, sm3, yogi, optimistic_gradient_descent, optimistic_adam, and optimistic_adam_v2. The novograd alias continues to use its internal decoupled approach.

Optimizers With Weight Decay

Historically designed with decoupled weight decay explicitly:

• adamw
• nadamw
• adamaxw

Already have weight decay added:

• adan
• lion
• lamb
• adadelta
• adafactor (named weight_decay_rate)

Newly updated to standard decoupled weight decay:

• adabelief
• adagrad
• amsgrad
• radam
• rmsprop
• sgd
• sm3
• yogi
• optimistic_gradient_descent
• optimistic_adam
• optimistic_adam_v2

Applying weight decay inside their specific transformation:

• novograd
• lars

Optimizers Without Weight Decay

Left without weight decay to serve as unmodified algorithms:
Kept for backward compatibility to match their "W" counterparts natively.

• adam
• nadam
• adamax

Modifies gradient signs (norm agnostic):
Weight decay could behave unpredictably since gradient magnitude is ignored.

• sign_sgd
• signum

Niche algorithms or raw-gradient wrappers:
Weight decay is not a standard practice for these methods.

• noisy_sgd
• rprop
• polyak_sgd
• lbfgs
• fromage