[feat] quantized GEMM support in the cuBLASLt and hipBLASLt backends

[harz05]

Summary (WIP):

Exposes through sofieBLAS the GEMM configurations that quantized inference needs, so SOFIE's quantized path can go through sofieBLAS

• Layouts, descriptors and algorithms are keyed by element, compute and scale type instead of being fixed to float. The float API and its behaviour are unchanged.
• matmul overload for int8 x int8 -> int32 accumulation (integer alpha/beta, optional strided batch).
• addOperationConfig overload that reports whether the library has a kernel for a typed configuration instead of ending the process, for callers with a fallback.
• Epilogue::Relu and DataType { F32, I8, I32 } enum in core.hpp; the matching rows in both Api tables.
• runGpuInt8Tests: NN, TN and batched int8 matmul with exact int32 comparison, for both tags.

[harz05]

Suggested change

	return ins->second.h;
	return &ins->second.h;