FP8 block scaling (blockwise, E4M3 with per-block scale)

Block scaling is a *separate recipe* from per-tensor FP8, and the TensorRT-LLM hardware matrix treats it as such: 'FP8(block scaling)' is Y on Blackwell sm100/103 and Rubin and is explicitly NOT Y on Ada Lovelace, while 'FP8(per tensor)' is Y on Ada (quant-trtllm-quantization). So blockwise FP8 is Blackwell-and-newer only in that stack, even though blockwise scaling as a numerical idea is not new. The doc's own footnote that Blackwell uses the MXFP8 scale recipe while SM90 uses FP32 scales is the clearest statement I found that 'blockwise' is not one format but a family. group_size 32 here is the MX block width, which is where the 32 comes from; other blockwise implementations use different widths, so this field is implementation-specific. No quality delta recorded.