Update the llama.cpp submodule to b10405 and port two API changes

[bogdanr]

The submodule moves from b10200 to b10405. Two functions in the sampler API change in this range. The Rust wrappers now agree with them.

llama_sampler_init_penalties has a new first parameter, n_vocab. Upstream moves n_vocab from llama_sampler_data into the penalty sampler (ggml-org/llama.cpp#26520). LlamaSampler::penalties has the same new first parameter. This is a breaking change, and each caller must add n_vocab as the first argument. The mirostat and logit_bias wrappers already take n_vocab as the first parameter, so the penalties wrapper is now consistent with them.

llama_sampler_init_dry no longer has the n_ctx_train parameter. Upstream removes the full-context window from the samplers that keep a history (ggml-org/llama.cpp#26524). LlamaSampler::dry keeps its public signature. The wrapper no longer reads n_ctx_train from the model.

The same upstream change also changes the meaning of penalty_last_n. A negative value selected the context length. Upstream now sets a negative value to 0, and 0 disables the penalty. The documentation of LlamaSampler::penalties shows the new behaviour.

The reason for this update is the output of the DeepSeek-V4 models. The current submodule gives incorrect text for these models. A build from b10405 gives correct text for the same model, the same prompt and the same hardware.

cargo build --workspace, cargo clippy and cargo test --features sampler are clean with the new submodule.

One note on CI: cargo fmt --check already fails on main in examples/llguidance.rs, llama-cpp-2/src/llguidance_sampler.rs and llama-cpp-2/src/model.rs. This branch does not add to that list, and it does not correct it, because the files are not related to this change. Tell me if you prefer a separate pull request for the format.