Document phi4-mini latency on additional hardware in docs/benchmark.md

[azurebeard]

Context

docs/benchmark.md currently lists indicative latency for phi4-mini and a few alternative models, but the numbers assume 8 vCPU on commodity hardware. Operators sizing a deployment on smaller VMs (2 vCPU, 1 vCPU + GPU, ARM, Raspberry Pi 5, etc.) don't have datapoints to anchor against.

Scope

Run the benchmark recipe at the bottom of docs/benchmark.md § "Repeatable benchmark recipe" against a hardware shape not already documented (ideally one of):

• 2 vCPU CPU-only VM
• 8 vCPU + consumer-grade GPU (e.g. T4, RTX 3060)
• ARM64 VM (AWS Graviton, Azure Ampere, or a Raspberry Pi 5)
• Apple Silicon (M1 / M2 / M3) — note that this is for development only; production target is Linux

Record p50 and p95 classifier latency over a 10,000-event run, plus the steady-state cache hit rate.

Acceptance criteria

• Extend the "Choosing a model" table in docs/benchmark.md with one new column (or one new row, depending on what's clearer) capturing the new hardware's numbers.
• Add a short note on any practical gotchas (e.g. "Apple Silicon — Ollama runs fine; pyspark and python-evtx wheels not yet ARM-native") if relevant.

Why this is a good first issue

No code. Pure data collection plus a small docs PR. Useful for any operator sizing their first deployment.