Discard transparent background pixels in the fragment shader instead of a separate CPU pre-processing step

[rdw-software]

This reduces the loading times, at the cost of having to perform more checks on the GPU. I highly doubt GPU usage will be an issue however, whereas loading times are very painful in general.

The drawback here is that if a single texture was to be reused many times, it would incur a runtime cost for each instance, whereas pre-processing has a startup cost shared between all instances. But I doubt it's relevant here, and there's no mechanism to reuse resources currently, anyway.

---

Before this change:

• Loading yuno takes about 2.1 seconds
• Loading icecastle takes about 2.2 seconds
• Loading pay_dun00 takes about 1.1 seconds
• Loading schg_dun01 takes about 2.4 seconds
• Loading prontera takes about 0.9 seconds

After this change:

• Loading yuno takes about 1.5 seconds
• Loading icecastle takes about 1.9 seconds
• Loading pay_dun00 takes about 1.1 seconds
• Loading schg_dun01 takes about 2.2 seconds
• Loading prontera takes about 0.9 seconds

The GPU frame time might be slightly higher, but I haven't measured since GPU usage is fairly low anyway.