runtimes: strided (pitched) buffer I/O + centralized stride helpers

[ariawisp]

Scope

• Adds row‑pitched (strided) buffer I/O on WGPU (read/write) for inner‑contiguous tensors with rank ≥ 2; contiguous path unchanged.
• CPU parity: pitched read/write with the same constraints; contiguous path unchanged.
• Centralizes stride logic in cubecl_runtime::stride and adopts it across CPU/HIP/CUDA/STD; removes duplicate helpers in STD.
• Adds client‑side size preflight and a clear error for data/shape mismatches.

Key Changes

• Stride helpers: add InnerContiguousRows, is_inner_contiguous_rows, and row_pitch_elems; keep contiguous_strides and is_contiguous.
• Allocation: for AllocationKind::Optimized and rank > 1, WGPU/CPU/HIP/CUDA allocate pitched rows (row pitch aligned to the backend’s alignment); Contiguous unchanged.
• Gating: WGPU/HIP/CUDA/CPU can_read_tensor accept fully contiguous or inner‑contiguous rows.
• Client: read and write preflight for stride compatibility; write_async/write helpers; to_client_tensor picks preferred allocation; size preflight returns IoError::InvalidSize.
• Utilities: expose pitched_rows_layout(..) and preferred_allocation_kind(..) for consistent planning.

Limits

• Only inner‑most contiguous, rank ≥ 2 layouts are supported; others return UnsupportedStrides.

Notes

• Prefer cubecl_runtime::stride::{contiguous_strides, is_contiguous, is_inner_contiguous_rows, row_pitch_elems}.
• cubecl_std::tensor::{is_contiguous, compact_strides} were removed.
• New error: IoError::InvalidSize { expected, actual } for create‑with‑data mismatches.

PR has been validated with Burn — no compilation or test errors.

[wingertge]

The code in general looks extremely clean, honestly one of the best PRs I've seen in a while. The describe function also seems extremely useful for inferring the correct variant of LinearLayout.

General comment, but I wonder if preferred_strides should be a device property. I haven't benchmarked the wgpu copy, but it seems like it might be slower compared to CUDA where the striding is handled by specialized GPU hardware and incurs no meaningful performance penalty. That way, a caller could decide on the allocation type based on their needs (i.e. optimized for long-lived tensors like weights, but the runtime preferred value for immediate inputs like images for a recognition model).

[ariawisp]

Implemented device-level preference for allocation layout:

• Added default_alloc_rank_gt1: AllocationKind to DeviceProperties.
  • WGPU/CPU default to Contiguous.
  • CUDA/HIP default to Optimized (pitched rows).
• ComputeClient::empty_tensor now picks the device default for rank > 1.
• ComputeClient::to_client_tensor uses the device default when the source is inner‑contiguous rows; otherwise falls back to contiguous.
• Centralized stride helpers remain the same (is_contiguous, is_inner_contiguous_rows, row_pitch_elems, pitched_rows_layout); IO still only accepts contiguous or inner‑contiguous rows.
• Keeps behavior focused: no changes to pitched copy logic; just device‑aware defaults. We can refine defaults with benchmarks later if needed.
• Workspace build and clippy are green.

[nathanielsimard]

@ariawisp The merge conflicts would need to be solved.

[ariawisp]

Branch has been rebased on main (cf40b4e) and confirmed CI workflow passes on my fork.