TensorRT-LLM docs/source/installation/build-from-source.md

The vendor's own build guide, read 2026-10-04 from the main branch. Establishes WHAT the build depends on, which is the half of the question that survives even when no duration is published: git-lfs checkout of large model assets; a purpose-built NGC 'devel' container (not a generic PyTorch image); specific Docker flags with stated reasons -- '--ipc=host is required to avoid Bus error (core dumped)', '--ulimit memlock=-1', '--ulimit stack=67108864'; a Python virtual environment created by the build script itself; Conan for C++ dependencies with a documented CMakeUserPresets.json and a relocatable CONAN_HOME; and a FetchContent third-party cache. Build-shape knobs, each of which is a direct admission that build time is a real cost: '-a "90-real" Build only for a specific GPU architecture (e.g. Hopper). Reduces compile time significantly'; '--fast_build Skip compiling some kernels to speed up compilation -- for development only'; '-j <N>' defaulting to all available CPUs; '--use_ccache'. Repetition conditions are also documented: 'All checkouts must stay on a commit whose C++ sources match the shared build, since the reused artifacts are not rebuilt.' AND the escape hatch when you only need Python changes: TRTLLM_USE_PRECOMPILED=1 pip install -e .