feat(cuda): add resilient native and grouped GEMM dispatch

[ppodolsky]

Summary

This PR adds resilient native GEMM paths to CubeCL's CUDA runtime:

• reclaim unused cached pages and retry once after BufferTooBig
• add backend-optional single and grouped GEMM descriptors with capability flags
• implement asynchronous BF16 GEMM with FP32 accumulation through cuBLASLt
• implement heterogeneous grouped GEMM through cublasGemmGroupedBatchedEx on CUDA 12.5+
• validate layouts, dimensions, batch strides, binding ranges, overlap, zero-K behavior, and stream ownership before launch
• cache cuBLASLt descriptors and algorithms by shape, with separate workspaces per CUDA stream

The allocation retry was previously proposed separately in #1438. It is consolidated here so the CUDA accelerator integration has one review thread.

Behavior and compatibility

Both GEMM primitives are optional runtime capabilities. Their server trait methods have default fail-loud implementations, so other backends require no changes. Callers can independently inspect features.matmul.accelerated_gemm and features.matmul.accelerated_grouped_gemm before dispatch.

CUDA advertises single BF16 GEMM only on SM80+. Grouped BF16 GEMM additionally requires CUDA 12.5. The grouped descriptor supports independent M, N, K, transpose state, and binding offsets per problem. Host metadata uses a bounded pinned staging ring guarded by CUDA events, so repeated and multi-stream launches remain asynchronous without overwriting in-flight arguments.

The allocator cleanup path runs only after storage reports BufferTooBig. It rechecks the selected pool after cleanup before attempting one new device allocation.

Dependency validation

• CubeK PR feat(attention): emit forward LSE and harden matmul tuning cubek#428 pins this PR head and compiles against it.
• Burn PR feat(cubecl): add native GEMM autotune candidate burn#5190 pins both dependency PR heads.
• Upstream main is the direct base of this branch.
• Downstream Cargo manifests reference only tracel-ai/* repositories.

Validation

• cargo fmt --all -- --check
• allocator retry regression test
• cargo check -p cubecl-runtime
• cargo check -p cubecl-cuda --examples
• A100 checks for padding, offsets, all transpose pairs, regular and broadcast batches, heterogeneous grouped shapes, repeated queued launches, asynchronous error propagation, and multi-stream ordering
• downstream CUDA linear forward and dx/dW/db parity under fusion, autotune, and autodiff
• 10,000 queued single GEMMs followed by one drain, confirming no per-call host synchronization

Performance

On an A100-SXM4 40GB, BF16 native single-GEMM dispatch beat or matched PyTorch matmul on 14 of 17 transformer-training shapes. A heterogeneous grouped expert projection measured 0.290 ms versus 0.438 ms for the prior padded batched path; activation time is excluded from both measurements.

[nathanielsimard]

Agreed that grouped GEMM is a gap in the cubecl/cubek/burn stack. That said, our goal is to avoid depending on pre-built binaries like cuBLAS wherever we can.

To that end, we're working on a tile API that should simplify cross-platform compatibility, cover all dtypes including quantized tensors, and compose with Burn's automatic fusion (both on-read and on-write).

We're happy to keep this PR around as a baseline to benchmark other implementations against, but we're trying hard to avoid the combinatorial explosion of kernels per platform × dtype × fusion group.