Rollup merge of #147686 - vrtgs:non-zero-isolate, r=joboet

update isolate_highest_one for NonZero<T>

## Rationale
Let `x = self` and
`m = (((1 as $Int) << (<$Int>::BITS - 1)).wrapping_shr(self.leading_zeros()))`
Then the previous code computed `NonZero::new_unchecked(x & m)`.
Since `m` has exactly one bit set (the most significant 1-bit of `x`), `(x & m) == m`.
Therefore, the masking step was redundant.

The shift is safe and does not need wrapping because:
* `self.leading_zeros() < $Int::BITS` because `self` is non-zero.
* The result of `unchecked_shr` is non-zero, satisfying the `NonZero` invariant. if wrapping happens we would be violating `NonZero` invariants.

why this micro optimization?
the old code was suboptimal it duplicated `$Int`’s isolate_highest_one logic instead of delegating to it. Since the type already wraps `$Int`, either delegation should be used for clarity or, if keeping a custom implementation, it should be optimized as above.

Author: matthiaskrgr