Bump llama.cpp to 4b98ab805 (latest master, +50 from b10830)

[nyo16]

Summary

Updates the vendored llama.cpp binding to the latest upstream master.

• Submodule vendor/llama.cpp: 465e49b9 (b10830) → 4b98ab805 — latest master, +50 commits
• Makefile LLAMA_COMMIT: bumped to 4b98ab805… to match (the pin used for Hex source-build clones)

The on-disk submodule tree had drifted to a stale 61881b1f (Aug 5); it is now aligned to the new pin.

API compatibility

Diffed all six headers the NIF includes (llama.h, ggml-backend.h, ggml-rpc.h, chat.h, json-schema-to-grammar.h, speculative.h) across 465e49b9..4b98ab805.

• Only change: llama_sampler_chain_n return type int → int32_t — identical on all supported platforms, no NIF impact.
• No llama_model_params / llama_context_params / llama_batch struct or default-params changes.

Verification

• Source build: LLAMA_BACKEND=metal mix compile --force — full llama.cpp rebuild to 100%, llama_nif.cpp compiled and linked clean.
• NIF load: backend_init/backend_free initialize Metal against the new lib.
• Model-backed smoke: mix test --include smoke --include embeddings on Qwen3.5-0.8B-Q4_K_M (gen) + Qwen3-Embedding-0.6B-f16 (emb) — 548 passed, 29 excluded. Exercised
  generation/chat/grammar/server + embedding paths.

Excluded: mtp / mtp_sidecar / mtp_cancel (no MTP model available) and rpc_live (no RPC worker).

Not included

• No @version bump or CHANGELOG.md entry — this is the binding update only, not a release.
• The three documented upstream defect workarounds (release-guide §2) were not re-audited.