Add transforms: grad, value_and_grad, vjp, jvp, vmap, compile, checkpoint

[sydneyrenee]

Summary

Implements all MLX function transforms, crossing the N-API boundary with JS function wrapping.

Function transformers (return new callable functions):

• grad(fn, argnums?) - compute gradients of scalar-valued functions
• value_and_grad(fn, argnums?) - returns both loss value and gradients
• vmap(fn, in_axes?, out_axes?) - vectorize over batch dimensions
• compile(fn, shapeless?) - JIT compile computation graphs
• checkpoint(fn) - recompute intermediates during backprop

Immediate transforms (compute and return results):

• vjp(fn, primals, cotangents) - vector-Jacobian product
• jvp(fn, primals, tangents) - Jacobian-vector product

Global flags:

• enable_compile() / disable_compile()

Architecture

JS functions are wrapped as C++ std::function via persistent Napi references, passed to MLX transform APIs (grad, compile, etc.), and the resulting transformed functions are wrapped back as callable JS functions. Closure data is stored with GC-safe cleanup via Napi::External attached to the function object.

Issues addressed

• Closes Implement mlx.core.grad() #166 (grad)
• Closes Implement mlx.core.value_and_grad() #167 (value_and_grad)
• Closes Implement mlx.core.vjp() #168 (vjp)
• Closes Implement mlx.core.jvp() #169 (jvp)
• Closes Implement mlx.core.vmap() #170 (vmap)
• Closes Implement mlx.core.compile() #171 (compile)
• Closes Implement mlx.core.enable_compile() #172 (enable_compile)
• Closes Implement mlx.core.disable_compile() #173 (disable_compile)
• Closes Implement mlx.core.checkpoint() #174 (checkpoint)

Test plan

• 9 new tests in batch6-transforms.test.ts, all passing
• Full suite: 239 passing, 9 pending (pre-existing)
• grad of x^2 returns 2x (single and multi-element)
• value_and_grad returns both loss and gradient
• Multi-arg grad with argnums selection
• vjp and jvp produce correct derivatives
• compile and checkpoint wrap functions correctly