docs: define staged CPU-to-GPU validation workflow

[MaharshPatelX]

Summary

• define implementation-ready, hardware-validated, and experiment-complete as separate project states
• document which portions of Parts 5-16 can be built safely before GPU validation
• add an eight-stage validation ladder from schema checks through full multi-seed runs
• require device-neutral code, CPU fixtures, synthetic embeddings, and resumable checkpoints
• record the immediate order: complete Part 5, stabilize Part 6 interfaces, validate the RX 7900 XTX, then continue
• update the tracker maintenance rules so GPU-dependent work cannot be marked Done from CPU tests alone

Why

The project can prepare substantial code before the GPU arrives, but it should not create one giant unvalidated implementation. This policy allows CPU-first progress while preserving explicit hardware and experiment completion gates.

Verification

• ruff format --check .
• ruff check .
• mypy src tests
• pytest -q -p no:cacheprovider (45 passed)
• git diff --check