llama.cpp v0.4.0

## Overview llama.cpp 0.4.0 adds initial Qwen3.8-Flash-Next and Nemotron-3-Puzzle support, on-demand tensor reading, per-slot server context limits, video input options, and a ggml update to 0.23.0 with major sparse flash attention and RDMA work. ### API changes - Added and ([#27794](https://github.com/ggml-org/llama.cpp/pull/27794)). - Added to quantize params ([#27795](https://github.com/ggml-org/llama.cpp/pull/27795)). - Bumped session/state versions for KV-cell token tracking ([#27762](https://github.com/ggml-org/llama.cpp/pull/27762)). - Added and ([#28250](https://github.com/ggml-org/llama.cpp/pull/28250)). - Made const-correct ([#28310](https://github.com/ggml-org/llama.cpp/pull/28310)). - Made const-correct ([#28307](https://github.com/ggml-org/llama.cpp/pull/28307)). - Added mtmd helper video/init options ([#24318](https://github.com/ggml-org/llama.cpp/pull/24318)). - Server multimodal helpers accept ([#24318](https://github.com/ggml-org/llama.cpp/pull/24318)). ### New models - Added initial Qwen3.8-Flash-Next () architecture support; optimization improvements are still pending ([#27742](https://github.com/ggml-org/llama.cpp/pull/27742)). - Added NVIDIA Nemotron-3-Puzzle-75B-A9B support ([#25444](https://github.com/ggml-org/llama.cpp/pull/25444)). - Added DSpark support for Nemotron 3.5 ([#27804](https://github.com/ggml-org/llama.cpp/pull/27804)). - Added support for nanbeige4.2-3B ([#27730](https://github.com/ggml-org/llama.cpp/pull/27730)). ### Core changes - Added lazy tensor reading ([#27794](https://github.com/ggml-org/llama.cpp/pull/27794)). - Renamed option to ([#27969](https://github.com/ggml-org/llama.cpp/pull/27969)). - Added ([#26622](https://github.com/ggml-org/llama.cpp/pull/26622)). - Added per-layer expert routing/FFN ([#28323](https://github.com/ggml-org/llama.cpp/pull/28323)). - Added KV-cell token tracking ([#27762](https://github.com/ggml-org/llama.cpp/pull/27762)). - Added n-gram history…