Modular package: break train_gpt_cpu.py into pgolf/ modules

[tayeva]

Description

Refactor train_gpt_cpu.py into importable, testable modules:

• pgolf/config.py — Hyperparameters dataclass
• pgolf/data.py — TokenStream, data loading
• pgolf/model.py — GPT, Block, Attention, MLP, RMSNorm, Rotary
• pgolf/optim.py — Muon optimizer, LR scheduling
• pgolf/quantize.py — int8 quantization + zlib
• pgolf/eval.py — BPB evaluation, tokenizer LUTs
• train_cpu.py — Thin CLI entrypoint

Acceptance Criteria

• All modules independently importable
• train_cpu.py reproduces exact same behavior as train_gpt_cpu.py
• Unit tests for each module
• ruff + mypy pass