axisMap generation

The recursion follows the structure of Shape.CanBroadcastTo:

• expand_dims inserts a new outer axis in the output, so we prepend 0.
• dim_eq / dim_1_to_n align an output axis with an input axis, so we prepend 1 (maps to input axis 0) and shift all tail mappings by +1 (because tail input axes are one level deeper).

def Runtime.Autograd.Cuda.Broadcast.shiftInputAxes (m : Array ℕ) : Array ℕ

def Runtime.Autograd.Cuda.Broadcast.axisMap {s₁ s₂ : Spec.Shape} : s₁.CanBroadcastTo s₂ → Array ℕ

Generate the CUDA axisMap array from a Shape.CanBroadcastTo proof.

def Runtime.Autograd.Cuda.Broadcast.broadcastArgs {s₁ s₂ : Spec.Shape} (cb : s₁.CanBroadcastTo s₂) : Array ℕ × Array ℕ × Array ℕ

Convenience bundle for CUDA broadcast kernels: (inDims, outDims, axisMap).