feat(cuda): add resilient cuBLASLt GEMM dispatch

[ppodolsky]

Summary

This draft consolidates the two CubeCL changes required by the downstream
accelerator stack into one upstream PR:

• reclaim unused cached pages and retry once when a device allocation reports
  BufferTooBig
• add a backend-optional accelerated-GEMM descriptor and capability flag at the
  server/client boundary
• implement asynchronous CUDA BF16 GEMM with FP32 accumulation through
  cuBLASLt
• validate layouts, dimensions, batch strides, binding ranges, overlap,
  zero-K behavior, and execution-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. Keeping it
here reflects the actual downstream dependency revision and gives CubeCL one
review thread for this accelerator integration.

Behavior and compatibility

The GEMM primitive is optional, not a global dispatch policy. Its trait method
has a default fail-loud implementation, so other backends need no changes.
Consumers check features.matmul.accelerated_gemm; Burn's existing autotuner
then compares native GEMM against its fused CubeK candidate per shape.

CUDA advertises BF16 only on SM80+. The implementation accepts row-major
matrices or transposed row-major views with strided batches and input
broadcast. Enqueueing remains asynchronous on the CubeCL stream; validation or
launch failures enter the existing stream error queue.

The allocator cleanup path runs only after the underlying storage reports
BufferTooBig. It rechecks the selected pool after cleanup before attempting
one new device allocation, leaving successful allocations and public APIs
unchanged.

Downstream validation required by the PR template

• CubeK PR feat(attention): emit forward LSE and validate CubeCL cubek#428 pins this PR head and compiles against it.
• Burn PR feat(cubecl): add native GEMM autotune candidate burn#5190 pins this PR and the CubeK PR heads.
• No downstream compatibility fixes beyond dependency pins were required.

Validation

• cargo fmt --all -- --check
• cargo test -p cubecl-runtime allocation_oom_reclaims_cached_pages_and_retries
• cargo check -p cubecl-cuda --examples on a non-CUDA host
• A100 cublas_gemm_check: padding, binding offsets, all transpose pairs,
  regular and broadcast batches, asynchronous error propagation, and
  three-stream producer/GEMM/consumer ordering
• downstream Burn CUDA linear forward and dx/dW/db parity under fusion,
  autotune, and autodiff
• 10,000 queued GEMMs followed by one drain confirmed no per-call host
  synchronization

Performance

On an A100-SXM4 40GB, BF16 native dispatch beat or matched PyTorch matmul on
14 of 17 transformer-training shapes. In the downstream 115.8M-parameter
transformer workload, letting Burn's autotuner arbitrate native GEMM against fused
CubeK improved steady-state throughput by 15.6% at batch 16 and 14.8% at batch
20, with matching loss and gradient norms.

The PR remains a draft until the upstream GPU matrix has run and dependency
pins can move to merge revisions.