spec 009: implement spinquant_hotstart.py (baseline + R_a-only modes)

Two new files in the openai#1736 submission dir:

spinquant_hotstart.py (~360 LOC):
- Imports from train_gpt.py for Hyperparameters/GPT/serialize/deserialize/
  eval_val/eval_val_ttt_phased/BatchedTTTLoRA/etc.
- Modes: baseline, internal_only (R_a only, per-layer per-KV-group, d_head
  rotation on V-output and O-input).
- full, port_1695 are stubs — raise NotImplementedError with explanation.
- Pipeline: load FP state_dict from HOTSTART_FP_CKPT -> apply rotations
  in-place on banked qo_bank/kv_bank -> optional pre-quant diagnostic eval
  -> call serialize() (GPTQ+compress) -> deserialize() -> quantized eval
  -> phased TTT eval -> write final.json.
- Reproduces the TTT eval block from train_and_eval (lines 2997-3075) in
  _run_ttt_eval() rather than refactoring the source file.

test_rotation_invariance.py (~250 LOC):
- CPU-only, standalone (no train_gpt.py import due to flash_attn_3/triton
  module-level deps).
- Self-contained minimal attention forward: Q/K/V projection from the
  banked tensors, RMSNorm on Q and K (matches real model's bound on
  attention logits; without this, trained weights saturate softmax and
  float noise in V amplifies catastrophically).
- Tests baseline (bit-exact identity) and internal_only (rel tolerance
  1e-4) against either synthetic random weights or spec 008's
  final_model.pt. Both pass cleanly (rel_max ~1e-6 on real checkpoint).
- Can load either banked (qo_bank/kv_bank) or unbanked
  (blocks.N.attn.*.weight) state_dict format.

Spec 009 updated: reduced scope to 2 modes (baseline, internal_only) for
this session; full and port_1695 deferred. Rationale in the spec: MLP
LeakyReLU-squared breaks R_m float-invariance, resid_mix can't be cleanly
folded through RMSNorm, both needing design before implementation.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: leon2k2k2k