Use cast_to for fused AsType in compiled Metal kernels

[katlun-lgtm]

Follow-up to #4224 (#4205). The compiled/fused path still flushes float32 and bfloat16 subnormals to False on .astype(mx.bool_): build_kernel in mlx/backend/metal/compiled.cpp emits a raw static_cast for is_static_cast primitives, so a fused AsType never reaches cast_to. This swaps it for cast_to, which is already in every generated kernel's preamble via metal::utils().

Repro — needs at least two fusable ops; a single-op mx.compile(lambda x: x.astype(mx.bool_)) doesn't fuse and falls back to the already-correct copy kernel:

import numpy as np, mlx.core as mx

sub = mx.array(np.array([0x00000001] * 4, dtype=np.uint32)).view(mx.float32)
print(np.array(sub.astype(mx.bool_)))   # [True True True True] -- fixed by #4224

f = mx.compile(lambda x: mx.broadcast_to(x, (2, 4)).astype(mx.bool_))
print(np.array(f(sub)))                 # all False on main     -- this PR

broadcast_to is deliberate: the fused broadcast preserves the input bits (0x00000001 in and out), so what fails is the cast, not FTZ arithmetic. Same failure on bfloat16. float16 is unaffected — Metal's static_cast<bool> on half is already correct (probed directly with mx.fast.metal_kernel).

Out of scope, and not claimed fixed: !=, >, == on subnormals still diverge from the CPU backend after this. Eager x + 0.0 zeroes the bits too, so that's ordinary Metal FTZ arithmetic, a different class from the cast. I also checked the other static_cast in this file (constant-input loads around line 121): it casts the printed constant to its own dtype, so there is nothing to fix there.

Tested on an M3 Max: the new test in test_compile.py fails on current main and passes with the change; full python/tests suite is green (801 passed, 51 skipped, 11293 subtests).

Investigation and test runs were AI-assisted; I've reviewed and understand the change.

Checklist

• 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)