Microscaling.jl

Array types for the OCP Microscaling Formats (MX):

• BlockscaledArray pairs elements of any type with per-block scales; compose with NarrowArray to pack sub-byte elements
• SwizzledArray presents storage rearranged by an Einops pattern at its logical shape; swizzle(x, :f8_4x128) — shorthand f8_4x128(x) — produces the 128×4 tiled scale layout Blackwell tensor cores read, zero-padding to whole tiles by default.