Fix quantized Gemma 3 MLP activation (gelu_pytorch_tanh, not SiLU)

[Sammy-Dabbas]

Fixes #3745

The quantized Gemma 3 MLP hardcoded SiLU, but Gemma 3 gates its feed-forward network with gelu_pytorch_tanh. The non-quantized gemma3.rs applies cfg.hidden_activation (gelu_pytorch_tanh for Gemma 3 checkpoints). The quantized model loads from GGUF metadata and has no config field, so it applies the activation directly, the same way quantized_glm4.rs hardcodes its Activation. This routes through the same code path as the non-quantized model (Activation::GeluPytorchTanh, the gelu tanh approximation).

SiLU and gelu_pytorch_tanh behave similarly, so the wrong one does not crash or produce obvious garbage; it just makes the output worse, which is why it went unnoticed. #3326 attempted a sliding-window plus GELU rework but was self-closed by its author, so this is the focused activation-only fix.

[jkantsios]

I was curious about the accuracy/performance impact this would have and produced the following perplexity score changes comparing SiLU and gelu_pytorch.

This definitely seems to confirm the efficacy of this change.