[WLDFA]

Summary

• add a direct binary producer path from ordinary computed CUBE_M32 tiles into TileArrayOutputRef slots
• preserve CUBE_M32 layout and emit the correct INIT/MIDDLE/LAST B.ASSEMBLE lifecycle
• keep parent capacity and per-fragment writer extent distinct without introducing B.SUBVIEW, TCVT, or TMOV materialization
• cover the binary elementwise family and document the supported API

Tests

• compiled the issue [TileOP API] Support binary assembly from ordinary CubeM32 fragments #288 [32,1] x 8 -> [32,8] reproduction to LLVM IR and object code
• verified eight producers, offsets 0 through 7, INIT + six MIDDLE + LAST, parent SizeCode 4, and writer SizeCode 1
• instantiated all 12 binary producer overloads
• verified expected diagnostics for layout, dtype, physical-shape, and valid-shape mismatches
• compiled existing SubTileView binary assembly, mixed binary, RMSNorm B128, and TileArray TCVT fixtures
• verify_target_cxx_frontend.sh passed the new issue [TileOP API] Support binary assembly from ordinary CubeM32 fragments #288 checks, then reached the pre-existing unrelated failure FAIL: Shared handle materialized to memory

Fixes #288