Fix SFB shared memory layout major selection

[migarci2]

smem_atom_layoutSFB currently selects its strides from majorSFA, so both mixed-major configurations get the opposite SFB ordering. This changes the predicate to majorSFB.

The host regression covers all four (majorSFA, majorSFB) combinations and compares the SFB shared-memory atom directly with tile_atom_to_shape_SFB. That gives a compile/pass check for (K, MN) and catches the otherwise masked (MN, K) mismatch at compile time.

Testing:

• compiled and ran CuTe_core.BlockwiseScaleLayoutSFBUsesOwnMajor with GCC 10.2.1
• confirmed the same test fails to compile on current main at exactly the two mixed-major assertions
• confirmed all four combinations compile with this change

Fixes #3535

[migarci2]

Friendly bump — ready for another look when you have a moment. Happy to address any feedback.

(Also have a few related small fixes open if useful: #3569, #3570, #3571, #3577.)