Making RCEIL the default for MXFP scale derivation

[frsun-nvda]

Propose to make RCEIL the default for MXFP scale derivation.

RCEIL is targeting to optimize the MXFP E8 scale factor w.r.t the ratio of block_amax / dtype_max. An example formula is: X_E8 = to_e8(ceil(log2(block_amax / dtype_max))). Real implementation is achieved through bit manipulation. Nvidia Blackwell architecture has instruction cvt.rp.satfinite.ue8m0x2.f32. TransformerEngine and CUTLASS have already adopted.
In contract, the OCP spec's ceil formula is X_E8 = to_e8(ceil(log2(block_amax)) - floor(log2(dtype_max))).
With block_amax = 150 and dtype_max = float8_e4m3.max = 448, RCEIL X_E8 = to_e8(ceil(log2(150 / 448))) = 126 while OCP ceil X_E8 = to_e8(ceil(log2(150)) - floor(log2(448))) = 127.

RCEIL is used to generate the cast-only slide for GTC 2025 to showcase direct-cast with MXFP8 could achieve iso-accuracy as BF16. We have also verified the model training and accuracy across various model sizes and token counts to ensure MXFP8 with RCEIL could achieve iso-accuracy with BF16.

The benefits of using RCEIL as the default is:

1. Users could enjoy the hardware acceleration for MXFP scale derivation on Blackwell architecture.
2. Users could enjoy cast-only MXFP8 out of the box with torchao to achieve iso-accuracy as BF16 for deployment.
3. Users could enjoy iso-accuracy MXFP8 training and downstream task as BF16.