Support runtime GEMM dimensions for dynamic shapes

[harz05]

Implements #10
sofieBLAS registers cuBLASLt matrix layouts at construct size (addLayoutConfig) and the matmul path looks them up by the runtime (rows, cols). With dynamic shapes the two disagree i.e. a Session constructed at one size and inferred at another misses the layout and throws std::out_of_range, so one Session only works at a single size.

This PR resolves each matrix's layout at matmul time instead: keep one cublasLtMatrixLayout_t per role (A/B/C) and stamp the runtime dims into it via cublasLtMatrixLayoutSetAttribute before each multiply. The descriptor is host-side metadata consumed by cublasLtMatmul at the call, so reusing one object across shapes is safe. addLayoutConfig becomes a no-op; the algorithm cache is unchanged (keyed by shape, re-queried per new size).

Approaches tried

Two ways to handle runtime dimensions were tried and benchmarked on ParticleNet onnx model, running one Session and sweeping infer over n_pf = 100, 80, 60, 40, 20 on a T4. The first was register-on-miss: create an immutable layout per distinct (rows, cols) on a miss and cache it in the map. The second was mutate-in-place (this PR): keep one layout per role and re-stamp the runtime dims via SetAttribute. Both give correct results (a construct-at-A / infer-at-B GEMM matches a CPU reference to ~1e-5), and they perform very close. mutate-in-place is marginally faster and a bit lighter on host memory.