[BUG] Complex vjps are wrong for cos, sqrt, and rsqrt (missing conjugate)

Describe the bug

mx.cos, mx.sqrt, and mx.rsqrt return the wrong gradient for complex inputs. cos's vjp delegates to its jvp, and sqrt/rsqrt form the derivative without conjugating it, so all three return f'(z) * cotangent instead of conj(f'(z)) * cotangent (the convention used by mx.exp/mx.log).

To Reproduce

import mlx.core as mx
z = mx.array(0.7 + 0.5j, mx.complex64)
g = mx.array(1.0 + 1.0j, mx.complex64)
print(mx.vjp(mx.cos, [z], [g])[1][0])
print(g * mx.conj(-mx.sin(z)))   # expected

Expected behavior

vjp == cotangent * conj(f'(z)), consistent with the other unary ops and finite differences.

Desktop

• OS: macOS
• Version: main (0.32.0.dev)

Start from the mx.cos, mx.sqrt, and mx.rsqrt autodiff entry points and compare their VJP/JVP behavior with mx.exp and mx.log. Reproduce the complex-input case from the issue, then verify that each VJP matches cotangent multiplied by the conjugate derivative and agrees with finite differences.