[FEA] Use native int8/uint8 vectors with int tensor cores for GPiPNN build #3059

GPiPNN initial build converts every datatype to fp16 to keep everything on a single execution path and using fp GEMMs. Directly using int8 and uint8 vectors when give should improve performance and reduce memory footprint for these datasets. This would give the following benefits:

• No conversion copy and require only 1 byte per element instead of 3.
• Half the operand traffic of the fp16 code path
• Potentially faster GEMM with int8 tensor cores
• Reduced shared memory footprint, making the fused path apply for larger dimension int8 datasets.