fix: prevent segfault when input exceeds batch size

[vlasky]

Summary

• Fix buffer overflow segfault when input tokenizes to more than 512 tokens (the hardcoded batch capacity)
• Size batch dynamically to actual token count instead of fixed 512
• Add bounds check with actionable error message when token count exceeds model's context size
• Fix memory leak: free tokens array after use

Details

The batch was initialized with a fixed capacity of 512 tokens, but the loop populating it had no bounds check. When an input tokenized to more than 512 tokens, this caused a buffer overflow and segmentation fault.

Resolves #20

Test plan

• Verify extension builds successfully
• Test with inputs that tokenize to more than 512 tokens - should return error instead of crashing
• Existing tests continue to pass