docs(v4-plan): mark ROI 1/2/3.A/3.5.A/4/6 done as Path A foundation ports

Status snapshot after this session's pipeline runs (Implementer -> Code Review
-> Perf Optimization -> Regression Tests, one per ROI):

DONE (all as minimal-edit copies of upstream PyTorch references, all with
numerical parity tests):
- ROI 1 V4MoE (sqrtsoftplus + aux-loss-free bias) — math-equivalent, 18 tests
- ROI 2 SequentialMTPHead (V3 architecture inherited in V4) — 16 tests
- ROI 3.A GDN Path A (FLA naive.py port) — 19 tests, parity vs PyTorch FLA
- ROI 3.5.A KDA Path A (FLA kda/naive.py port) — 19 tests, parity vs PyTorch FLA
- ROI 4 mHC primitives (TileKernels torch/mhc.py port) — 12 tests, 5 parity tests
- ROI 6 Engram primitives (TileKernels torch/engram.py port) — 7 tests, 3 parity tests

DEFERRED (Metal/TileLang/CUDA kernel work, ~1-3 days each):
- ROI 3.B-F GDN Paths B/C/D/E + dispatch
- ROI 3.5.B-D KDA Paths B/C/D
- ROI 3.7 head-to-head benchmark + receipt (needs Paths B/C/D/E first)
- ROI 5 FlashMLA absorb (no Python ref in upstream)
- ROI 7 Lightning Indexer (act_quant/fp8_index/rotate_activation deps)
- ROI 8/9 NSA / CSA+HCA research spikes

Plugin invariant intact: zero modifications under cppmega_mlx/. All new code
lives in cppmega_v4/ side-by-side. The parallel TileLang/TVM/TVM-FFI perf
work on the existing model is unaffected.

91/91 v4 tests + 147/147 combined regression green at HEAD.

Author: apstenku123