fix: clamp mingru hidden state before i16 cast

[Sal-ami]

The Q48 accumulator y in mingru can exceed the i16 range after the >> 33 shift. On inputs that push the hidden state high enough, (y >> 33) as i16 wraps silently in release mode, producing incorrect hidden state values with no diagnostic.

This adds a one-line clamp before the cast. The gate g is already clamped to [0, 16777215] on the line above, so the analogous treatment for the output is overdue.

In practice the trained weights probably keep things in range, but "probably" is doing a lot of work for something that affects every forward pass. A silent correctness issue is worse than a crash.

This is a one-line fix. The change is the smallest thing that prevents the bug. No tests because the repo doesn't have a test infrastructure yet and adding a framework for one clamp assertion would be the opposite of lazy. The maintainer can decide if this warrants a targeted regression test.