Pass the target as the MTP draft's ctx_other so Gemma 4 sidecars load (#91)

[nyo16]

Fixes #91.

Root cause

llama.cpp's gemma4-assistant head builds its draft context over the target's: the draft's llama_kv_cache aliases the target's cell vector (other->v_cells_impl) and the graph reads the target's
tok_embd. Its constructor throws Gemma4Assistant requires ctx_other to be set when llama_context_params.ctx_other is null (llama-context.cpp, LLM_ARCH_GEMMA4_ASSISTANT).

context_create built every context from llama_context_default_params() and never exposed the field. The Qwen 3.8 sidecar path from v0.8.44 worked because qwen35 never reads the pointer; a Gemma 4 pair
could not get past Context.create.

Reproduced pre-fix with Unsloth gemma-4-E4B-it-Q4_K_M.gguf + MTP/mtp-gemma-4-E4B-it-Q8_0.gguf on Metal: validate_head/2 passes (head n_layer_nextn == 4, n_embd_out 2560 on both), then
llama_init_from_model fails with exactly the error in the issue.

Fix

Mirrors upstream's common_speculative_init_result, which sets cparams.ctx_other = ctx_tgt for every draft context.

• c_src/llama_cpp_ex/llama_nif.cpp — context_create takes std::optional<fine::ResourcePtr<LlamaContext>> ctx_other and forwards it as params.ctx_other.
• c_src/llama_cpp_ex/llama_nif.h — LlamaContext holds the peer's ResourcePtr, so the target cannot be freed under a draft whose KV cells alias it.
• LlamaCppEx.Context.create/2 — new :ctx_other option (%Context{} or nil), added to the structural keys and documented.
• LlamaCppEx.MTP.init/2 — passes the target as the draft's :ctx_other unconditionally. llama-context.cpp only reads the field for GEMMA4_ASSISTANT / EAGLE3 / DFLASH, so the in-file and Qwen
  sidecar paths are unchanged.
• test/mtp_model_test.exs — timing_us reports a ckpt bucket asserted ckpt > 0, which pins Qwen 3.8's hybrid-SSM snapshotting rather than the binding. It now branches on
  TestModels.seq_rm_kind(:mtp): :full targets must be billed for snapshots, everything else must not — the same condition as the NIF's needs_ckpt gate.

Creating a Gemma 4 draft context without a target still fails. That is llama.cpp's contract; Context.create/2 documents it.

Verification (Apple M4 Max, Metal source build)

Run	Result
Gemma 4 E4B pair, pre-fix	{:error, "failed to create context"} / Gemma4Assistant requires ctx_other to be set
Gemma 4 E4B pair, post-fix	MTP.init ok; greedy " Paris.", 53% acceptance at n_draft: 3
--include mtp_sidecar on the Gemma pair	6/6, incl. greedy equivalence with plain target decode
--include mtp (Qwen3.5-0.8B-MTP-Q8_0, in-file head)	19 passed
--include smoke --include embeddings	548 passed, 29 excluded
mix test (no model)	428 passed, 149 excluded
mix format --check-formatted, mix credo --strict	clean

Not run: the Qwen 3.8 27B sidecar pair (no local copy). Unaffected by construction — qwen35 ignores ctx_other — and the identical code path passed on the Qwen 0.8B in-file head.

Docs

• CHANGELOG.md — ## Unreleased → Fixed entry
• LlamaCppEx.MTP moduledoc — the two sidecar head architectures and what each needs from the target
• README.md — Gemma 4 E4B pair added to the MTP model list
• docs/release-guide.md — Gemma mtp_sidecar run added to the test matrix, since a Qwen pair alone cannot exercise the ctx_other wiring