v2.3.2 update made some large models stop working due to Out Of Memory

[felladrin]

For the record, the v2.3.2 update (#179) made some large models (2.3 GB+ [Gemma 3 4B, Qwen 3 4B and Llama 3.1 Nemotron Nano 4b, all at Q4_K_S with 4096 context]) stop working due to Out Of Memory.

Not sure if it's because llama.cpp started requiring more memory for those models.

Current workaround options:

• Reduce the context size
• Downgrade to Q3 quant
• Downgrade the @wllama/wllama to v2.3.1