Effects of removing swizzles, MI-200

[krzysz00]

swizzle-perf.zip shows the perf impacts (on a MI-200 node) of disabling the swizzle code.

The overall notes:

• In general, this makes things faster
• Untuned performance on some configs (certain f32 gemms, some NHWC convolutions) is worse
• Tuned performance is generally better

So, we can probably turn off swizzles for an overall performance gain, especially since we're not as worried about MI-100 anymore. (they weren't generally a significant perf win on MI-100 in any case)