CI: Add CUDA 13.4 support - #10294

Mirrors cupy#10152 for CUDA 13.4:

- schema.yaml: register CUDA 13.4, cuda-python 13.4, and add 13.4 under
  NCCL 2.30 / cuTENSOR 2.4 / cuSPARSELt 0.9.1
- matrix.yaml, config.pbtxt: add cupy.linux.cuda134{,.multi} + cupy.win.cuda134
  as the new '@Push mini' target; demote 13.3 to '@nightly full'
- matrix.yaml: bump cupy.linux.cuda13x-cuda-python to CUDA 13.4 + cuda-python 13.4
- ci/versions.yml: build.version 13.3.1 -> 13.4.1
- windows/_flexci.ps1: add 13.4 branch, point 13.x fallback to 13.4
- docs/source/install.rst: append v13.4 to the CUDA Toolkit list

The regenerated Dockerfiles / .sh / .gitattributes / config.tags.json /
coverage.rst come from '.pfnci/generate.py'.

CUDA gained Windows-on-Arm host support in CTK 13.4, so wire the
GHA build lane to also produce cupy-cuda13x win_arm64 wheels alongside
the existing linux-64 / linux-aarch64 / win-64 lanes.

Follows the NVIDIA/cuda-python#2789 pattern:

- New `build-win-arm64` job in ci.yml, gated on
  `startsWith(CUDA_BUILD_VER, '13.')` so pre-13.4 backport branches
  cleanly skip it.
- Uses only `CUDA_BUILD_VER` (there is no CUDA 12 x arm64 pairing).
- Python 3.10 is excluded on `win-arm64` -- CPython upstream has no
  official 3.10 Windows ARM64 build (nuget-cpython and
  actions/setup-python's manifest both drop it), so the matrix
  entry is unbuildable.
- Runner: `windows-11-arm` (GitHub-hosted).
- `fetch_ctk` action learns the `windows-arm64` redist subdir;
  supporting Python tools (`env-vars`, `prepare_wheel_build.py`,
  `wheel_configs.py`, `fetch_ctk_redistrib.py`) get the
  matching platform entries.
- cuTENSOR preload and rapidsai/sccache are skipped for `win-arm64`
  because NVIDIA has not yet shipped WoA cuTENSOR binaries and
  rapidsai/sccache has no WoA build. Both are marked in-code as
  temporary; re-enable when the upstreams catch up.
- No test/import step for `win-arm64` -- no GHA runner has an
  ARM64 GPU. This lane is build-only for now.

The nvidia/cuda registry ships 13.4.1-* only; the initial 13.4.0
image was never published, so the auto-generated Dockerfiles
failed to pull `nvidia/cuda:13.4.0-devel-ubuntu22.04`.

Bump `.pfnci/schema.yaml`'s 13.4 full_version to 13.4.1 (matching
the precedent set by 12.9 -> 12.9.1) and regenerate the affected
Dockerfiles via `.pfnci/generate.py`.

NVIDIA's apt repo has no 'libnccl2=2.30.*+cuda13.4' build -- the
+cuda13.4 line jumps straight to NCCL 2.31.2. cuda134 / cuda134.multi
/ cuda13x-cuda-python were failing apt install with

  E: Version '2.30.*+cuda13.4' for 'libnccl2' was not found

Move those matrix rows to NCCL 2.31, keep the existing 13.3 lane on
2.30 (which does ship 2.30.7-1+cuda13.3), and regenerate the affected
Dockerfiles via .pfnci/generate.py.

Two spots gated on 'host-platform == win-64' need to cover win-arm64
too, since both target Windows:

- build-wheel.yml: the LongPathsEnabled registry check is OS-wide,
  not arch-specific -- WoA needs the same guard.
- prepare_wheel_build.py: Windows uses ';' as PATH separator on
  every arch, so win-arm64 was picking ':' and generating malformed
  CUPY_INCLUDE_PATH / CUPY_LIBRARY_PATH values.

Both cases now match on startsWith('win-'). Other 'win-64' uses in
the tree (runner map, sccache URL triple, PFN CI artifact name)
remain intentionally per-arch.

Matches the .pfnci/matrix.yaml + schema.yaml addition for the CUDA
13.4 lanes.