Add Instruction Fine-tuning Support for LoRA with Assistant-Only Loss

[makaveli10]

This PR adds support for instruction fine-tuning using LoRA adapters with efficient masked loss computation.

Key Features:

• Assistant-Only Loss: --assistant-loss-only flag enables training only on assistant responses, ignoring system/user prompts in conversation datasets
• Masked Cross-Entropy Loss: CPU and VUlkan shader implementation for computing loss only on assistant tokens
• Count-Equal Op: CPU and Vulkan-shader for masked token counting for accuracy metrics during training
• Chat ML Template Support: Native support for conversation-format datasets with automatic ChatML rendering
• Custom Chat Templates: Support for Jinja chat templates via --chat-template custom-templat.jinja to render JSON conversation datasets (compatible with HuggingFace apply_chat_template)

This enables efficient instruction fine-tuning of language models using LoRA adapters with proper loss masking for conversational AI datasets/applications.

This PR is built on top of #43 so, we should land that first.

[zoq]

LR scheduler sample comands:

constant (default):
./build/bin/llama-finetune-lora -m Qwen3_0.6B.Q8_0.gguf -f trump.txt -ngl 999 -c 256 -b 256 -ub 256 --flash-attn off

cosine:
./build/bin/llama-finetune-lora -m Qwen3_0.6B.Q8_0.gguf -f trump.txt -ngl 999 -c 256 -b 256 -ub 256 --flash-attn off --learning-rate 3e-4 --lr-scheduler cosine --lr-min 1e-5 --weight-decay 0.01

linear:
./build/bin/llama-finetune-lora -m Qwen3_0.6B.Q8_0.gguf -f trump.txt -ngl 999 -c 256 -b 256 -ub 256 --flash-attn off --lr-scheduler linear --lr-min 1e-5

[zoq]

Added --warmup-ratio to match HF training parameters, if warmup-steps is used it overwrites --warmup-ratio.