[Docs] Document CUDA_ARCH_NAME=Manual and CUDA_ARCH_BIN in runtime FAQ

[mahmoudshehata74]

What

Expands the answer to the no kernel image is available for execution on the device
FAQ entry with:

• CUDA_ARCH_NAME=Manual + CUDA_ARCH_BIN, which is currently undocumented
• Compute capability values for recent architectures (Ampere 86, Ada 89, Blackwell 120)
• A command to verify which architectures ended up in the built library

Why

The current answer suggests CUDA_ARCH_NAME=All or named architectures
(Turing, Volta, Pascal). All builds every supported architecture and adds
significant build time, and the named examples predate Ampere.

CUDA_ARCH_NAME=Manual with CUDA_ARCH_BIN lets users target exactly the
architectures they need, but it appears neither in this FAQ nor in the CMake
option tables of the source-compile guides.

The ordering caveat matters in practice: if CUDA_ARCH_NAME=Manual does not
precede CUDA_ARCH_BIN, the architecture list is silently ignored and the build
completes without the requested kernels — a failure only discovered at runtime.

The cuobjdump verification turns a multi-hour failure into a one-minute check.

Context

Encountered while building the C++ inference library for an RTX 5060 (sm_120).
Verified on Ubuntu 24.04, CUDA 13.0, Paddle v3.3.0 — the resulting library
contains both sm_89 and sm_120.

Related discussion: PaddlePaddle/Paddle#79725

[CLAassistant]

All committers have signed the CLA.