docs: memory footprint & host-link streaming analysis

[booth-algo]

Adds doc/memory-footprint-and-streaming.md — reasoning notes on:

• Footprint: SmolVLM2-256M at MXFP8 ≈ 288 MB weights + tens-of-MB KV → fits in 512 MB onboard DDR (the "~1 GB" GPU figure is CUDA overhead PLENA doesn't have).
• Implication: the HostStream / LayerSwapping models (lib/memory/src/streaming.rs) are for weights-exceed-DDR; for this model use plain resident DDR (no streaming).
• Instructions: load the program into DDR once; don't stream it per-cycle.
• DDR3 vs host link: bandwidth table + the latency point (local DDR3 ~15–50 ns vs PCIe ~0.5–1 µs), and the transport-vs-memory framing. For the actual boards: Nexys local DDR3 (~1–3.2 GB/s) ≫ USB2 (~38 MB/s, no PCIe); V80 HBM (819 GB/s) ≫ PCIe.
• Real bottleneck: DDR bandwidth (~3–4 tok/s on Nexys), not the host link.
• Validation: run LayerSwapping with capacity = 512 MB and confirm zero swaps; note the emulator currently models DDR as 128 GiB so the limit isn't enforced today.

Captured from a design discussion; no code changes.

[booth-algo]

Consolidated into #58 (combined WIP reference PR) — the memory-footprint doc now lives on kev/profile-and-opt.