fix(wgpu): clamp compute dispatch cube count to u16::MAX (NVIDIA Vulkan)

[thewtex]

Summary

A small batch of independent correctness fixes in cubecl-wgpu, plus the same
rank-0 stride bug fixed in two sibling crates that carry a copy of the helper.
Each fix is its own commit so they can be reviewed.

Fixes

1. Clamp dispatch cube count to the enforced per-dimension limit (cubecl-wgpu)

Some adapters — notably the Vulkan backend on NVIDIA hardware — report a
max_compute_workgroups_per_dimension larger than the limit the driver
actually validates against at dispatch time. create_server fed that
adapter-reported value straight into HardwareProperties::max_cube_count, so
the launch logic treated an oversized [N, 1, 1] cube count as legal and let
it reach wgpu, which then rejected the dispatch with a validation error once
N exceeded the real cap of u16::MAX (65_535).

This bites any single large elementwise dispatch — e.g. a 52,428,800-element
launch needs 52_428_800 / 256 = 204_800 workgroups along X, well past 65_535.

The fix clamps the reported value to u16::MAX via a new clamp_max_cube_count
helper, keeping it consistent with the cap already advertised by
<WgpuRuntime as Runtime>::max_cube_count.

2. Detect 8-bit storage via the 8-bit feature struct (cubecl-wgpu)

In register_types, storage8 was computed from ext_feat.buf_16 and its
uniform_and_storage_buffer16_bit_access field — a copy-paste of the
storage16 line above it. As a result 8-bit storage capability tracked the
16-bit device feature, so I8/U8 types could be marked storable based on the
wrong feature. Now reads ext_feat.buf_8.uniform_and_storage_buffer8_bit_access.

3. Name AUTO_GRAPHICS_BACKEND in its invalid-value error (cubecl-wgpu)

The error printed for an unrecognised backend reads the AUTO_GRAPHICS_BACKEND
env var but told the user to check GRAPHICS_BACKEND, sending them to the wrong
variable. Corrected the message.

4. Guard contiguous_strides against rank-0 shapes (cubecl-wgpu, cubecl-cpu, cubecl-runtime)

contiguous_strides iterated (0..rank - 1), which underflows and panics when
rank == 0 (a scalar / rank-0 shape). Changed to (0..rank.saturating_sub(1))
so a rank-0 shape yields empty strides. The same helper is duplicated in
cubecl-cpu and cubecl-runtime, fixed in all three.

Notes:

• cubecl-std's contiguous_strides uses a different, subtraction-free
  implementation and is not affected.
• The pitched-allocation path in ContiguousMemoryLayoutPolicy
  (cubecl-runtime) already guards its rank - 2 loop behind rank > 2, so
  only the standalone helper needed the fix.

Testing

• New unit tests for clamp_max_cube_count (caps inflated limits, preserves
  honest ones, agrees with WgpuRuntime::max_cube_count) and for
  contiguous_strides (rank-0 returns empty, row-major correctness) in each of
  the three crates.
• cargo test, cargo clippy --all-targets, and cargo fmt --check pass for
  the touched crates; cubecl-wgpu additionally checked under --features spirv
  to compile the Vulkan path.