vt CPU attention: the per-element dtype dispatch is 64.6% of the kernel, and 219 call sites share the defect

Row: VT-CPU-ELEM-DISPATCH

vt::AttentionCross and vt::Attention spend most of their CPU time resolving
an element type and a byte offset, not doing arithmetic.

Both CPU kernels (src/vt/cpu/cpu_ops.cpp) read every operand element through
LoadF32, which switches on t.dtype and multiplies the offset by
vt::SizeOf(t.dtype) once per element. SizeOf was defined out of line in
src/vt/dtype.cpp and the build enables no LTO — CMakeLists.txt sets no
INTERPROCEDURAL_OPTIMIZATION and passes no -flto — so it was a cross-
translation-unit call per element. AttentionCrossKernel's score loop calls
LoadF32 twice per element, around a body that is one multiply and one add.

A perf profile of vt::AttentionCross alone, at the LTX-2.5 connector's
video shape (Tq=S=1024, H=32, D=128, f32), on this x86-64 AVX-512 devbox:

64.6% of the kernel's own CPU time resolves an element type and an address; the
arithmetic and the softmax are the 15.60% line. The kernel runs at ~3.6 GFLOP/s
on 20 threads.

This is not an LTX-2.5 problem. LoadF32/StoreF32 appear at 219 call sites
across 64 kernels in src/vt/cpu/cpu_ops.cpp and src/vt/cpu/cpu_paged_attn.cpp;
attention is only where it was measured. Every CPU model path pays it.

The repair is the transformation MatmulOneChunk already applies against
MatmulOneChunkRef in the same file: resolve the element type once, outside the
loops, and walk typed pointers. It touches no output's accumulation order, so it
is bit-exact rather than close.

Measured on dgx-class hardware by the predecessor row
(.agents/specs/ltx25-connector-gemm.md, ## Owed), and reproduced here.

Start in src/vt/cpu/cpu_ops.cpp and src/vt/cpu/cpu_paged_attn.cpp, using MatmulOneChunk and MatmulOneChunkRef as the stated precedent. Audit the 219 LoadF32/StoreF32 call sites and the Attention CPU kernels, then verify that element types and offsets are no longer resolved per element while accumulation remains bit-exact.