Add 0-d NDArray support

[Copilot]

NDArray[A] previously rejected empty shapes, precluding 0-dimensional arrays needed for scalar-tensor bridge patterns (e.g., MathTrig[NDArray[Double]] where zero/one/fromDouble must produce shape-agnostic values that broadcast universally).

Core changes

• NDArrayCheck.scala: Remove the shape.length == 0 rejection from shapeCheck. 0-d arrays are now valid; the dimension loop simply doesn't execute.
• ndarray.scala: Add isScalar: Boolean property; add NDArray.scalar[A](value) factory producing shape=[], strides=[], numel=1.
• ndarrayOps.scala: Add scalar (0-arity read) and setScalar (0-arity write) extension methods.
• ndarrayDoubleOps.scala / ndarrayFloatOps.scala: Add 0-d guards to all binary ops, in-place ops, and comparison ops. A 0-d operand is auto-promoted by delegating to the existing scalar extension — this is the only implicit broadcast relaxation; all other shape mismatches still require explicit broadcastTo/broadcastPair.

Semantics (NumPy-aligned)

val s = NDArray.scalar(3.0)                  // shape=[], ndim=0, numel=1
val v = NDArray(Array(1.0, 2.0, 3.0), Array(3))

s + v                                        // Array(4.0, 5.0, 6.0) — 0-d broadcasts to any shape
s - v                                        // Array(2.0, 1.0, 0.0)
NDArray.scalar(2.0) + NDArray.scalar(3.0)    // scalar(5.0) — 0-d + 0-d = 0-d

val a = v.reshape(Array())                   // 1-element 1-d → 0-d
a.scalar                                     // 1.0
a.setScalar(99.0)

NDArray.zeros[Double](Array())               // scalar(0.0)
NDArray.ones[Double](Array())                // scalar(1.0)

All existing broadcast machinery (broadcastShape, broadcastTo, broadcastPair) already handles 0-d correctly without modification.