fix: convert weight scale shape[N, K/128] to [ceil(N/128), ceil(K/128)]

[JiaoliangYu]

Motivation

per_block_fp8 online quantization in ATOM can produce degenerate generation (all-f outputs or immediate EOS) because the weight-scale layout emitted by the online quant path does not match the layout expected by the existing linear FP8 blockscale GEMM backend.

Technical Details

In LinearBase.online_quantize_weight for online_quant_type == QuantType.per_1x128, we no longer
use the generic per-group quant output directly (weight_scale shaped like [N, K/128]). Instead,
we explicitly quantize weights into 128x128 blockscale format and emit:

• q_weight: shape [N, K] (same logical weight shape),
• weight_scale: shape [ceil(N/128), ceil(K/128)].

This aligns online-quantized weights with the current linear compute path
(gemm_a8w8_blockscale*), which consumes 128x128 block scales.

Additional fix included:
Added float8_e4m3fnuz compatibility in online quant dtype checks and avoided redundant normalize
marking when quant output is already fnuz (relevant on gfx942).

Test Plan

Test Result

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.