Fix gluon attention example for fp8

[masahi]

Currently, running python python/examples/gluon/01-attention-forward.py fails with

python/examples/gluon/01-attention-forward.py:386:14: error: source and result must have the same logical storage size (4096 vs 2048)
    p_tmem = s_tmem._reinterpret(config.dtype, [config.SPLIT_M, 2 * config.BLOCK_N], config.p_tmem_layout)

This happens when the benchmark is run with fp8. pytest python/examples/gluon/01-attention-forward.py doesn't fail simply because it doesn't test on fp8.

The regression is from #10243 but the problematic code hardcoding a fixed column count like config.BLOCK_N // 2 has been there for a long time. Previously it only worked due to the loose convention on reinterpret.