server: off-by-one in deferred_create_final_checkpoint + shed unused build + AGENTS.md memory budget docs

Three changes from the QA pass:

1. server-context.cpp: deferred_create_final_checkpoint off-by-one.
   The previous fix in 1585f98b3 sourced prompt_n_tokens from
   slot.prompt.n_tokens() - slot.n_decoded. That recovers the prompt
   boundary in most cases but loses a token on the k-th generation:
   at the moment the deferred flag fires, prompt.tokens has N + (k-1)
   tokens (the in-loop push_back of the k-th sampled token has not yet
   happened -- handle_last_sampled_token runs after the sampling loop,
   not inside it), n_decoded is k, and the right answer is
   N = (N + k - 1) - k + 1. Without the +1, pos_max covered [0, N-2]
   instead of [0, N-1] -- the last prompt token's KV entry was stripped
   from the checkpoint, the next cold start reprocessed that 1 token,
   and f_keep/f_sim metrics were off by 1.

2. Shed unused tools and the embedded UI from the default CachyLLama
   server build. CMakeLists.txt and tools/CMakeLists.txt drop the
   app/ unified binary, examples/, batched-bench, gguf-split,
   completion, perplexity, parser, tts, cvector-generator,
   export-lora, fit-params, and results. tools/server/server-http.cpp
   and tools/server/CMakeLists.txt gate the embedded-frontend assets
   behind LLAMA_BUILD_UI. LLAMA_BUILD_EXAMPLES/APP/UI now default OFF.
   Server, imatrix, llama-bench, quantize, tokenize, mtmd, and rpc
   are still built. Keeps the build small and focused on the server
   path while letting LLAMA_BUILD_UI=ON still get the web UI.

3. AGENTS.md context checkpoint section: the worst-case memory budget
   paragraph said the 2 GiB was a cap (small checkpoint counts got
   capped at 2 GiB); _ckpt_memory_budget() actually returns
   std::max(2 GiB, n_ctx_checkpoints * 400 MiB), so 2 GiB is the
   floor -- small counts get the full 2 GiB working set, large counts
   scale up. Rewording reflects the actual std::max behavior.