ANE from julia (research)

[samtalki]

long-shot. Apple's Neural Engine (ANE) is exposed via CoreML only. is there
a practical julia path for FW LMO / gradient eval?

investigate:

• CoreML.jl wrapper — exists?
• direct Metal Performance Shaders Graph (MPSG) APIs that route to ANE
• bridge that compiles a julia op to a CoreML model on-the-fly

if any of these is reasonable, consider the matmul-heavy gradient and the
spectraplex eigen path. tracking only — not a near-term priority.