Bump llama.cpp to 2d97363 (b9870), release v0.8.32

[nyo16]

Summary

Updates the vendor/llama.cpp submodule from f708a5b2c (b9846) to 2d973636e (b9870) — 24 upstream commits — and releases v0.8.32.

API compatibility

No NIF changes were required. ggml/include/ggml.h, ggml/include/ggml-backend.h, common/common.h, common/chat.h,
common/json-schema-to-grammar.h, common/sampling.h, and common/speculative.h are all unchanged. The only touched header the binding compiles
against is include/llama.h, and its diff is purely additive: a new llama_ftype_name() helper and a new llama_model_ftype() getter (#25134) —
the binding calls neither.

Notable upstream changes

• llama API: add llama_model_ftype() / llama_ftype_name() for reading a model's quantization type (#25134)
• model: register t_layer_inp for qwen3next (#25141)
• chat: trim messages sent to the StepFun parser, fixing long reasoning loops (#25238)
• spec/dflash: support spec-draft-p-min in DFlash (#25246)
• CUDA: topk-moe fusion for 288 experts (#25267); remove redundant copies after gated_delta_net (#23940); __restrict__ + PDL for
  FlashAttention (#25185); fix KQ mask stride truncation/overflow in flash_attn_mask_to_KV_max (#24945); fix get_rows_back for >65535 rows
  (#25103); fix Gemma E4B MTP FlashAttention (#25148)
• ggml-cpu: AVX2 optimization for the nvfp4 dot product using a UE4M3 LUT (#23961)
• opencl: precompiled binary kernel loading (#23042); initial q1_0 support (#25160)
• hexagon: flash-attention rework (#25085)
• common/server: HF primary split as model path (#25194); bracketed IPv6 literals in URL authorities (#25140); SSE keepalive pings (#25241);
  cpp-httplib 0.49.0 (#25218)

Verification

• ✅ Clean rebuild from source (mix clean && mix compile)
• ✅ Full test suite: 216 passed, 1 skipped (generation: Qwen3.5-0.8B, embeddings: Qwen3-Embedding-0.6B)
• ✅ All 7 end-to-end smoke tests pass (generation, streaming, chat templates, JSON-schema grammar, raw GBNF, embeddings)
• ✅ mix format --check-formatted clean
• ✅ Dialyzer: 0 errors