gemm_fp8_blockwise consumes too much GPU memory

[kyle-256]

use Primus dev/kyle branch. Run
BACKEND=megatron EXP=examples/megatron/configs/llama3_8B-pretrain.yaml bash ./examples/run_pretrain.sh
the GPU memory will be occupied to 98%~99%, which needs to be optimized.
@xiaobochen-amd

[hann-wang]

Seems that activations are stored in high precision.
In deepseek-v3 tech report, activations are stored in fp8 and dequant-requant during backward.