[Bug]: Add numerical stability guard for SigLip2 logit_scale

🐛 Describe the bug

When calling SigLip2Model.forward() during training with an unconstrained logit_scale parameter that has grown to 100.0, the model produces NaN values in logits_per_image and logits_per_text instead of the expected similarity matrix with finite values.

🔄 Steps to Reproduce

import torch
from kornia.models.siglip2 import SigLip2Model, SigLip2Config

config = SigLip2Config()
model = SigLip2Model(config)

# 1.Simulate large logit_scale (can occur during training)
model.logit_scale.data.fill_(100.0)

pixel_values = torch.randn(2, 3, 224, 224)
input_ids = torch.randint(0, 1000, (2, 10))

# 2.Forward pass produces NaN without clamping
output = model(pixel_values=pixel_values, input_ids=input_ids)
print(output.logits_per_image)  # Contains NaN values

💻 Minimal Code Example

import torch
from kornia.models.siglip2 import SigLip2Model, SigLip2Config

# Create model
config = SigLip2Config()
model = SigLip2Model(config)

# Simulate extreme logit_scale value (can occur during training)
model.logit_scale.data.fill_(100.0)  # Large positive value

# Prepare inputs
pixel_values = torch.randn(2, 3, 224, 224)
input_ids = torch.randint(0, 1000, (2, 10))

# Forward pass
output = model(pixel_values=pixel_values, input_ids=input_ids)

# Check for NaN (occurs without clamping)
print(f"logit_scale value: {model.logit_scale.item()}")
print(f"Effective scale (exp): {output.logit_scale.item()}")  # Should overflow
print(f"Contains NaN: {torch.isnan(output.logits_per_image).any()}")
print(f"logits_per_image:\n{output.logits_per_image}")

✅ Expected behavior

The model should maintain numerical stability regardless of learned parameter values.

❌ Actual behavior

exp(100) overflows, producing inf/NaN in logits and loss.

🔧 Environment

- Kornia version: 0.8.2
- PyTorch version: 2.9.1
- Python version: 3.11.14
- OS: macOS (darwin)
- Installation method: pip / from source
- CUDA/cuDNN version: N/A (CPU testing)
- GPU model: N/A

📝 Additional context

The overflow can occur during:

• Extended training runs where logit_scale learns extreme values
• Transfer learning scenarios with aggressive learning rates
• Any situation where the learnable temperature parameter grows beyond safe numerical bounds

🤝 Contribution Intent

• I plan to submit a PR to fix this bug
• I'm reporting this bug but not planning to fix it