Optimize `planar_snake` calculation.
In typical usage, we expect `tile_width` to be a power-of-two. In this case, using a `cond`, rather than a `select`, is faster, as the cost of branches is not equal.
This code is ~10% faster when `tile_width` is a power-of-two, and also faster in most other cases:
```
BEFORE
================================================================================
Shape TileW Latency / us
--------------------------------
1024x1024 4 84.19
1024x1025 4 167.58
1025x1024 4 193.79
1025x1025 4 195.97
1024x1024 5 249.78
1024x1025 5 358.53
1025x1024 5 159.58
1025x1025 5 197.92
AFTER
================================================================================
Shape TileW Latency / us
--------------------------------
1024x1024 4 84.19
1024x1025 4 155.55
1025x1024 4 163.58
1025x1025 4 169.76
1024x1024 5 173.41
1024x1025 5 282.11
1025x1024 5 167.61
1025x1025 5 193.89
```
PiperOrigin-RevId: 969625172
