Bump llama.cpp to 52fb93a2b (30 commits)

[nyo16]

Summary

• Bumps vendor/llama.cpp from b28a2f372 → 52fb93a2b (30 upstream commits).
• No public API changes. Existing NIF and LlamaCppEx.MTP bindings continue to work unchanged.
• README/MTP docs unchanged — Metal MTP optimization (#23114) is still unmerged upstream, and the documented LlamaCppEx.MTP API
  surface is unaffected.

Headlines

MTP / speculative

• Backend sampling for the MTP draft path (#23287) — adds a backend_sampling field to common_params_speculative_draft (default
  true, additive). Our NIF doesn't touch this field, so behavior is unchanged.
• Skip logit computation via inp_out_ids (#23433) — internal optimization on the MTP draft path.
• Fix nullptr crash in common_speculative_get_devices_str (#23386).
• Server: free draft/MTP resources on slot sleep (#23461) — fixes an upstream VRAM leak (not user-facing for our binding, but good
  hygiene if the server code is ever reused).
• Doc typo fix (#23435).

Other notable upstream changes

• llama: null-buffer crash fix in llm_graph_input_attn_kv_iswa on SWA-only models (#23131).
• vocab: Carbon-3B HybridDNATokenizer support (#23410).
• server: re-inject subcommand when the router spawns children under the unified binary (#23442).
• app: introduce the llama unified executable (#23296); add batched-bench, fit-params, quantize, perplexity subcommands
  (#23459); show version (#23426).
• mtmd: merge HunyuanOCR into HunyuanVL + OCR vision precision fix (#23329); DeepSeek-OCR image-processing fixes +
  img_tool::resize padding refactor (#23345); fit_params accounts for mmproj
  (#21489); WAV MIME-type variants + audio format detection (#23396).
• ggml: correct iface-method check before 2D-get fallback (#23306).
• metal: optimize pad + cpy (#23354).
• CUDA: Programmatic Dependent Launch (PDL) for Hopper+ (#22522); RDNA3 Q6_K MMVQ nwarps tuning
  (#23349).
• vulkan: optimize IM2COL shader (#22685).
• opencl: refactor backend initialization (#23318).
• hexagon: ssm-conv fix for large prompts (#23307); HMX quantized matmul rework
  (#23368).
• snapdragon: toolchain v0.6 (#23369).
• webui: max image size option (#22849); reactive isMobile in viewport store
  (#23330); pointer-events fix on hidden div wrapper (#23390); text attachments
  before message content in chat-completions payload (#23406); improved UI dev git hooks
  (#23403).
• docker: copy conversion files (#23370).

Test plan

• NIF rebuilds cleanly against 52fb93a2b (forced full rebuild on macOS / Metal backend).
• mix test → 89 passed, 4 skipped.
• CI green on Linux + macOS workflows.
• Smoke-test LlamaCppEx.MTP.stream/3 against a Qwen 3.6 MTP GGUF locally (optional — internal MTP changes are additive but worth a sanity check before release).