Improve collection of performance/throughput metrics

[orionpapadakis]

This PR addresses #104 .

Introduces a RunMetrics singleton that accumulates timing data across the full inference stack and prints a unified performance summary, replacing the previous ad-hoc LastRunMetrics record.

Metrics collected

Core (aligned with Ollama-style reporting):

• load_duration — model file load time, measured in ModelLoader (both overloads)
• prompt_eval_count / prompt_eval_duration — prefill token count and wall-clock time
• eval_count / eval_duration — decode token count and wall-clock time
• total_duration — full inference wall-clock time

TornadoVM-specific:

• tornado_task_graph_compile_duration — plan/graph construction time
• tornado_task_graph_warmup_duration — JIT compilation (withPreCompilation())
• read_only_weights_copy_in_duration — first-execution weight upload (forceCopyInReadOnlyData())

All values stored in nanoseconds.

Derived throughput printed at end of run

• Standard engine: achieved tok/s: X. Tokens: N, seconds: T
• Prefill-decode engines: per-phase rates (prefill tok/s, decode tok/s) in addition to total
• TornadoVM init breakdown printed only when --verbose-init is set — no new flags introduced

---

How to test

All commands below use Llama-3.2-1B-Instruct-F16.gguf. Each should print a ==== Performance Metrics ==== block at the end.

Standard engine — CPU

./llama-tornado --model <model> --prompt "Tell me a joke"
Expected: single "achieved tok/s" line (no prefill/decode split)

Standard engine — GPU

./llama-tornado --gpu --ptx --model <model> --prompt "Tell me a joke"
Expected: single "achieved tok/s" line

Standard engine — GPU + --verbose-init

./llama-tornado --gpu --ptx --model <model> --prompt "Tell me a joke" --verbose-init
Expected: throughput line + "Compilation & CodeGen / Warmup / Read-only weights Copy-in" lines

Prefill-decode — CPU

./llama-tornado --model <model> --prompt "Tell me a joke" --with-prefill-decode
Expected: Total + ¬Prefill + ¬Decode tok/s lines

Prefill-decode — GPU

./llama-tornado --gpu --ptx --model <model> --prompt "Tell me a joke" --with-prefill-decode --verbose-init
Expected: per-phase throughput + TornadoVM init breakdown

Batch prefill-decode — GPU + CUDA graphs

./llama-tornado --gpu --ptx --model <model> --prompt "Tell me a joke" \ --with-prefill-decode --batch-prefill-size 32 --cuda-graphs --verbose-init
Expected: per-phase throughput + TornadoVM init breakdown