Tensor/window Codegen always emits `float *`, which is incorrect for custom memories

[yamaguchi1024]

• @instr(x : f32[...] @ AVX2) gets compiled into (..., float *x) instead of (..., __mm256 x)
• _data is correct, but the vanilla {dst} dictionary's C string is incorrect. It uses float * (or sth like that) as a member of window struct, instead of the respective hw memory
• Maybe we should make use of mem (not just prim_type and shape) in window codegen