feat(default): promote Phi-3.5-mini to recommended default model

[unamedkr]

Summary

Phi-3 architecture support landed in #65 and validated end-to-end as the best speed/quality combo we ship — vocab 32K + 3.8B params makes the lm_head matmul the fastest of any registered model. Promote Phi-3.5-mini to the default everywhere.

Default model progression:

1. Llama-3.2-1B (original)
2. SmolLM2-1.7B (feat(feedback): Quick Wins from 2026-04-12 external user report #59 — switched after vocab-size benchmark feedback)
3. Phi-3.5-mini (this PR)

Changes

Code

• _MODEL_REGISTRY reordered with Phi-3.5-mini first, with a comment block marking it as the default and explaining the reasoning
• cmd_chat_default (no-subcommand chat) now picks Phi-3.5-mini
• Module docstring + Model.from_pretrained example use Phi-3.5-mini
• CLI --help epilog: examples lead with `phi-3.5-mini`; backwards-compat block mentions `smollm2` / `llama3.2:1b` as alternatives

Docs

• README.md Quick Start: Phi-3.5-mini is the recommended default. CLI examples and Python `from_pretrained` example updated. Benchmark/perf sections still reference SmolLM2/Llama because those are historical measurement data.
• README.ko.md: same changes mirrored in Korean.
• bindings/python/README.md (PyPI README): replaced "Basic question answering" with "Quick start (auto-download)" using `from_pretrained`. Added a multi-turn chat example using `m.chat()` + KV cache reuse, and API reference entries for `Model.chat()` and `Model.from_pretrained()`.

Verified

• `ctest --test-dir build` → 35/35 passed
• Full build clean (no new warnings)
• Phi-3.5-mini end-to-end inference test still produces coherent multi-paragraph output:

  "Name three planets in the solar system."
  → "Sure, here are names of three planets from our Solar System: 1. Earth — Our home planet... 2. Mars — Known as the Red Planet... 3. Jupiter — The largest planet in our solar..."

• `available_models()` returns `['Llama-3.2-1B', 'Phi-3.5-mini', 'Qwen3.5-0.8B', 'SmolLM2-1.7B', 'SmolLM2-135M']`
• `MODEL_ALIASES['phi-3.5-mini']` and friends resolve correctly
• `cmd_chat_default` source confirms `args.model = "Phi-3.5-mini"`
• `quantcpp --help` epilog reflects the new defaults

Test plan

• Unit tests pass
• Phi-3.5-mini end-to-end inference (regression)
• CLI `--help` shows new examples
• `available_models()` includes Phi-3.5-mini
• Manual: `pip install -e .` from this branch + `quantcpp` (no args) downloads Phi-3.5-mini and starts chat

🤖 Generated with Claude Code