Add nested Swift Array initializers (and reverse) for MLXArray

[john-rocky]

Closes #161.

Adds row-major nested Swift Array initializers for MLXArray (and the matching reverse extraction methods) for 2-, 3-, and 4-dimensional arrays. Higher ranks are intentionally left out — that question was the main reason this issue had been waiting (cf. @davidkoski: "We can certainly do it for a fixed number of dimensions… the types will be tricky" for arbitrary N). 4-D covers `NCHW` / `NHWC` image literals without introducing a macro dependency, so it felt like a natural stopping point.

API additions

```swift
// 2-D
MLXArray([[1.0, 2.0], [3.0, 4.0]]) // .float32, shape [2, 2]
MLXArray([[1, 2], [3, 4]]) // .int32, shape [2, 2] (matches MLXArray([Int]))
MLXArray(int64: [[1, 2], [3, 4]]) // .int64
MLXArray(converting: [[1.5, 2.5], [3.5, 4.5]]) // .float32 from Double

// 3-D (same four shapes)
MLXArray([[[1.0], [2.0]], [[3.0], [4.0]]]) // .float32, shape [2, 2, 1]

// 4-D (same four shapes), useful for NCHW / NHWC literals
MLXArray([[
[[1.0, 2.0], [3.0, 4.0]],
[[5.0, 6.0], [7.0, 8.0]],
]]) // .float32, shape [1, 2, 2, 2]
```

Each initializer flattens the nested array and forwards to the existing `MLXArray(_: [T], _ shape:)` path, so the dtype rules match the 1-D initializers exactly (notably: `[Int]` → `.int32`, `int64:` produces `.int64`, `converting:` casts `Double` → `Float`). Uneven inner-array lengths trigger a `precondition` failure, matching the existing convention for shape-mismatch elsewhere in `MLXArray+Init.swift`.

Reverse extraction

@JimWallace asked in the issue for the inverse direction ("MLXArray to nested Swift arrays would also be handy"). Adding three matching helpers on MLXArray:

```swift
let m = MLXArray([[1.0, 2.0], [3.0, 4.0]])
m.asArray2(Float.self) // [[1.0, 2.0], [3.0, 4.0]]
m.reshaped([2, 2, 1]).asArray3(Float.self) // 3-D nested
m.reshaped([1, 2, 2, 1]).asArray4(Float.self) // 4-D nested
```

`asArray2` / `asArray3` / `asArray4` reuse the existing `asArray(_:) -> [T]` path, so cross-dtype extraction follows the same `asType` semantics — `asArray2(Float.self)` on an int32 array implicitly converts to float32. Rank mismatches trigger a `precondition` failure, matching the existing 1-D `asArray` style.

Files

• `Source/MLX/MLXArray+NestedArrayInit.swift` — initializers and reverse extraction (purely Swift, no Cmlx calls beyond the existing 1-D path).
• `Tests/MLXTests/MLXArray+NestedArrayInitTests.swift` — 12 tests covering dtype inference, shape, round-trip on plain 2/3/4-D, round-trip after `reshaped`, and cross-type extraction (`asArray2(Float.self)` from an int32 array).

Verification

```
swift build # Build complete!, no new warnings
```

`swift test` for the broader package hits the pre-existing "Failed to load the default metallib" error (#349 / #342) before any of the new tests can be exercised on this machine. The added code is pure Swift Array ↔ existing `MLXArray(_: [T], _ shape:)` plumbing, so the runtime path is identical to the 1-D initializers. I'd appreciate a maintainer running the new test class through Xcode to confirm; happy to iterate on the test set if helpful.

Intentional non-goals

• 5-D and higher: per the issue thread ("the types will be tricky" for arbitrary N), and because 5+D literals are uncommon in user code.
• Macro-based variant: @robertmsale floated an expression macro (`#mlx([[...]])`). That is a heavier dependency choice and orthogonal to this PR; can be added on top if desired.
• `throws` instead of `precondition`: I matched the precondition style already used by `MLXArray(: [Int], _ shape:)` and `asArray(:)`. Happy to convert to `throws` if you prefer that for the shape-mismatch / rank-mismatch paths.

[john-rocky]

Apologies — I missed @VDurocher's #392 when surveying open PRs and opened a duplicate. Their `MLXNestedArray` protocol approach supports arbitrary depth via the type system, which is a cleaner design than the fixed 2-D/3-D/4-D overloads I wrote here. Closing in favor of #392.

If any of the round-trip tests against `reshaped` here are useful for #392, please feel free to lift them — no attribution needed.

Sorry for the noise.