Fix kernel cache collision in Compiled constructor - #3427

The constructor moves inputs and constant_ids into members but the loop building kernel_lib_ continued to iterate the moved-from parameters. After the move those parameters are empty, so the kernel name omits the per-input dtype suffix and a kernel compiled for one input type can be incorrectly reused for another. With ops fused around the cast (e.g. chained astype feeding matmul or einsum), this surfaces as byte-level corruption of the output.

Reproducer:

import mlx.core as mx

def f(x_u16, y_i8):
    a = x_u16.astype(mx.int32)
    b = y_i8.astype(mx.int32)
    af = a.astype(mx.float32)
    bf = b.astype(mx.float32)
    return mx.matmul(bf.reshape(2, 1), af.reshape(1, 2))

x = mx.array([10, 20], dtype=mx.uint16)
y = mx.array([1, 2], dtype=mx.int8)
print(f(x, y))                 # [[10, 20], [20, 40]]
print(mx.compile(f)(x, y))     # [[5130, 10260], [0, 0]]  <-- wrong

5130 = 0x140A is the little-endian byte pattern of the two uint16 inputs packed into one int32, confirming the kernel was specialized for the wrong dtype.

Switching the loop to use the inputs_ / constant_ids_ members restores the dtype suffix and the bug goes away.

Verified locally against the issue reproducer, exhaustive dtype matrix, and fresh-process isolation tests (44/44 + 9/9 pass). I can add an inline regression test if you guys would like a specific format / location.

Fixes #3338.

Checklist

Put an x in the boxes that apply.

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (if needed)