Fix nvcc CUDA 12.0/12.1 C++20 ranges parse error (#3907)

[nlohmann]

• The changes are described in detail, both the what and why.
• An existing issue is referenced: error: expected initializer before ‘<’ token  #3907.
• The Code coverage remained at 100%. A test case for every new line of code.
• If applicable, the documentation is updated.
• The source code is amalgamated by running make amalgamate.

What and why

#3907 reports that nlohmann/json.hpp fails to compile under nvcc at -std=c++20
(error: expected initializer before '<' token in iteration_proxy.hpp's
enable_borrowed_range specialization), but compiles fine at C++17. The issue was
previously closed as "un-reproducible in CI": the existing ci_cuda_example job built
against the json-ci image's CUDA 11.0 toolkit at cuda_std_11, and CUDA 11 has no
C++20 support at all, so the job could never hit this code path. A separate job
(ci_nvhpc) and a previously closed PR (#4796) both use CUDA 12.9, a version where the
bug is already fixed — so neither of those could catch it either.

This PR was done in two stages, both on this branch:

1. Diagnose first. Turned ci_cuda_example into a matrix pulling official
nvidia/cuda:*-devel-ubuntu22.04 images directly (same pattern as ci_nvhpc) across 8
CUDA versions (11.8.0 through 12.6.3) at cuda_std_20, to empirically confirm which
versions are actually affected instead of trusting an unverified issue-thread comment.
Result, reproduced against real nvcc:

CUDA (nvcc)	Result
11.8.0	fails at CMake configure — no C++20 support in CUDA 11, unrelated to #3907
12.0.1	fails with the exact #3907 error
12.1.1	fails with the exact #3907 error
12.2.2 – 12.6.3	all pass

So the affected range is exactly nvcc 12.0.x/12.1.x, fixed by NVIDIA from 12.2 onward —
nothing broader or narrower.

2. Fix it. Added a targeted guard to the JSON_HAS_RANGES detection chain in
include/nlohmann/detail/macro_scope.hpp, disabling ranges support specifically for
nvcc 12.0/12.1 (detected via __CUDACC_VER_MAJOR__/__CUDACC_VER_MINOR__), matching the
style of the neighboring GCC-11/libstdc++ carve-outs already in that file. This is the
part that protects real client code — any project compiling this header with nvcc
12.0/12.1 at C++20 gets JSON_HAS_RANGES disabled automatically, regardless of their own
build system. Regenerated single_include/nlohmann/json.hpp accordingly.

Also broadened tests/cuda_example/json_cuda.cu (previously only called dump()/
erase()) to exercise operator==/operator<=> (three-way comparison, gated
independently by JSON_HAS_THREE_WAY_COMPARISON) and range-based iteration, so the CI
matrix gives real evidence about the fix's scope rather than a guess. This did catch a
real issue: once 11.8.0 started building at C++17 (see below), its use of <=>
unconditionally broke, since that operator isn't valid pre-C++20 syntax — fixed by
gating those two lines behind JSON_HAS_THREE_WAY_COMPARISON, the same way the library
itself does internally.

tests/cuda_example/CMakeLists.txt no longer hard-requires cuda_std_20 (which failed
outright on anything older than CUDA 12.0). It now picks the newest C++ standard the
detected nvcc version actually supports — 20 for CUDA ≥ 12.0, 17 for CUDA ≥ 11.0, else
11 — so older toolkits still build the smoke test at whatever standard they support
instead of erroring. This is local to our own test project only (a header can't control
its own -std= invocation); the actual protection for downstream users is the
macro_scope.hpp guard above.

With the fix in place, every CUDA version can build successfully at its own standard, so
the matrix no longer needs the full 8-version diagnostic sweep to avoid a permanently red
leg. Right-sized ci_cuda_example down to 3 legs, each testing something distinct:

• 11.8.0 — exercises the new C++17 fallback path end-to-end
• 12.1.1 — the permanent regression guard for error: expected initializer before ‘<’ token  #3907
• 12.6.3 — recent CUDA/C++20 coverage

Also updated the CUDA row in docs/mkdocs/docs/community/quality_assurance.md's compiler
table to match (previously listed a stale CUDA 11.0.221 / Ubuntu 20.04 entry).

Final CI on this branch: all three ci_cuda_example legs (11.8.0, 12.1.1, 12.6.3)
pass.

Changes

• include/nlohmann/detail/macro_scope.hpp / single_include/nlohmann/json.hpp: nvcc
  12.0/12.1 JSON_HAS_RANGES guard (the actual fix for error: expected initializer before ‘<’ token  #3907).
• tests/cuda_example/json_cuda.cu: broadened smoke test (comparisons + range-based
  iteration), <=> usage gated by JSON_HAS_THREE_WAY_COMPARISON.
• tests/cuda_example/CMakeLists.txt: graceful per-nvcc-version C++ standard selection
  instead of hard-requiring cuda_std_20.
• cmake/ci.cmake: dropped the stale -DCMAKE_CUDA_HOST_COMPILER=g++-8 pin on the
  ci_cuda_example target so it falls back to the stock host compiler in each image.
• .github/workflows/ubuntu.yml: ci_cuda_example now matrices over
  nvidia/cuda:{11.8.0,12.1.1,12.6.3}-devel-ubuntu22.04 directly instead of the pinned
  ghcr.io/nlohmann/json-ci image, with a lukka/get-cmake step added.
• docs/mkdocs/docs/community/quality_assurance.md: updated the CUDA compiler-version
  table row(s) to match.