Bump llama.cpp to b10665 (ca3d5a3e1)

[nyo16]

Summary

Bumps vendored llama.cpp b10582 (e85caa81e) → b10665 (ca3d5a3e1), 83 builds. Submodule and LLAMA_COMMIT in the Makefile
move together, per docs/release-guide.md.

Binding changes

One real API break: upstream replaced nlohmann::ordered_json with its own pimpl wrapper common_json (common/json.h) across
common/chat.h and common/json-schema-to-grammar.h.

• json_schema_to_grammar_nif now parses with common_json::parse; the direct <nlohmann/json.hpp> include is gone.
• nlohmann is still the backing parser behind the wrapper, so the NIF's size/depth bounds on untrusted schema text
  (recursive-descent stack-overflow guard) are unchanged.

Everything else in the range is additive, no NIF edits needed:

• include/llama.h: new llama_tensor_read_lazy field in llama_model_params — the NIF builds params from
  llama_model_default_params(), unaffected. Session/state-seq versions bump 9→10 / 2→3.
• common/speculative.h: only gained functions (common_speculative_n_max, synthetic-acceptance helpers); every
  common_speculative_* callsite is signature-identical.

Verification

macOS (Metal, M1 Max), rebuilt with LLAMA_BACKEND=metal:

Run	Result
Default suite (no model)	428 passed, 149 excluded, 0 failures
--include mtp_sidecar (Qwen3.8-27B-Q4_K_M + mtp-Qwen3.8-27B-Q4_0 sidecar head)	434 passed, 0 failures

mix compile --warnings-as-errors and mix format --check-formatted clean.