1.26.0

This release lets yzma run in a browser, and it checks what it installs.

• WebAssembly. llama.cpp becomes a WebAssembly module, and a Go program compiled
  by TinyGo drives it through the new pkg/llamawasm package. There are three builds,
  and the JavaScript glue takes the best one the browser can run: WebGPU, more threads,
  or one thread. Text, embeddings, and images all work, with chat templates and tool
  calling. The names and the order of the calls are the same as in pkg/llama and
  pkg/mtmd. See wasm/README.md.
• Digest checks. Install checks the SHA-256 of each file against the manifest that
  llama-cpp-builder publishes. A version takes the form VERSION@sha256:DIGEST to pin
  the manifest as well, and the new yzma verify command checks an installation later.
• A pinned llama.cpp release. This release of yzma installs llama.cpp v0.4.0 when
  no version is given. Use --version latest to get the most recent nightly build.
• mtmd gets TokenizeFromParts and InputPart, and llama gets a wrapper for
  llama_sampler_init_temp.
• The installer follows the ROCm asset names of llama.cpp from b10356 and b10767.

What's Changed

• llama, mtmd: follow the upstream llama.cpp b10655 ABI changes by @deadprogram in #298
• download: install the llama.cpp version pinned by this yzma release by @deadprogram in #300
• llama: follow the upstream llama.cpp b10675 ABI change by @deadprogram in #301
• loader: keep the FFI argument types alive by @deadprogram in #304
• llama: align lazy-loading API names with llama.cpp b10679 by @deadprogram in #305
• Run llama.cpp in a browser with TinyGo by @deadprogram in #306
• WebGPU in the browser by @deadprogram in #307
• A model with eyes in the browser by @deadprogram in #308
• download: follow the llama.cpp ROCm asset rename at b10356 by @deadprogram in #310
• wasm: start the threads of llama.cpp correctly by @deadprogram in #311
• wasm: add chat templates and tool calling by @deadprogram in #312
• docs: show the WebAssembly state of each function in the roadmap by @deadprogram in #313
• download: support llama.cpp ROCm 10.0 asset names from b10767 by @deadprogram in #316
• llamawasm: do not put the model in the browser cache by @deadprogram in #317
• wasm: run the WebGPU build in Firefox by @deadprogram in #318
• mtmd: add TokenizeFromParts and InputPart by @deadprogram in #320
• wasm: use the CPU in Firefox and document browser needs by @deadprogram in #324
• download: check the digest of each asset and add a verify command by @deadprogram in #323
• download: accept a SHA-256 digest with a version by @deadprogram in #326
• wasm: bind the batch and memory calls of ABI 6 by @deadprogram in #328
• download: publish and read the digest of the version manifest by @deadprogram in #330
• Release 1.26.0 by @deadprogram in #331
• download: do not read a digest mismatch as a file that is not there by @deadprogram in #332

Full Changelog: v1.25.0...v1.26.0