docs: add docs/fork/ feature guide + README/AGENTS pointers

[Piggidragon]

Establishes docs/fork/ as the home for docs specific to this maintained fork,
with the first page a user-facing guide to everything the branch adds over
upstream master.

Files

• docs/fork/README.md - the guide (browsing to docs/fork/ shows it).
• README.md - a 6-line note at the top linking to docs/fork/. llama/dev's
  README is otherwise byte-identical to upstream; nothing fork-specific was in
  it to remove.
• AGENTS.md - one line under "Useful Resources": docs/fork/ is where fork
  docs go, upstream docs stay in docs/, and the guide tracks user-facing flag
  changes.

What the guide covers

• New CLI flags - what each does, when to use it, how, and the LLAMA_ARG_*
  env var: --kv-cpu-pinned, --recurrent-state-offload, --kv-gpu-layers,
  --phase-aware-workspace, --live-context-workspace,
  --spec-draft-ubatch-size, --spec-draft-kv-gpu-layers, --spec-mtp-rs-planes.
• Optimizations that run with no flag - compact causal attention masks, the
  quantization-preserving host KV copy, shared target/MTP workspaces.
• llama-bench additions and the extra environment variables.
• Ready-to-use recipes (single GPU large context, hybrid model, two asymmetric GPUs).

Material from branches not yet on llama/dev, marked inline in the guide:

• ggml-meta : split a host-resident KV cache by head #57 - -as / --attn-split, the -sm tensor + --no-kv-offload
  correctness fix, the --kv-gpu-layers slowest-link-first ordering.
• sched: pipeline the delivery of a host-resident KV cache #39 - --kv-pipeline-depth / --kv-pipeline-budget, the llama-bench
-kvcp / -rso / -kvpd / -kvpb switches, GGML_SCHED_TRANSPORT_DEBUG.
• branch moe-cache-drafting - the CUDA MoE expert cache:
  --moe-expert-cache-size, --moe-expert-cache-l2-pinned-mb,
  --experimental-logs, plus the automatic grouped-decode / prefetch / bias
  residency behaviour and the layer-split-only and speculative caveats.

Notes for reviewers

• Docs only, no code change. Style follows docs/multi-gpu.md, ASCII only per
  AGENTS.md. All internal links checked.
• #39 (docs/kv-transport-pipelining.md, docs/repro/) and
  moe-cache-drafting (README feature section, commit eb7f74e) carry their own
  docs. When they rebase/land, those should move into docs/fork/ and the
  README section should collapse to the pointer added here, so there is a single
  source.
• Draft while ggml-meta : split a host-resident KV cache by head #57, sched: pipeline the delivery of a host-resident KV cache #39 and moe-cache-drafting are unmerged; the flag names and
  defaults for that material match those branches as of now.

AI usage disclosure: drafted with Claude Code, see the commit trailers.

🤖 Generated with Claude Code

https://claude.ai/code/session_013Bs926e6SdsxnkrqqKq5HP