fn decomp_tx: Do not zero txa before initialization

[rinon]

This adds some complexity, but it improved performance significantly, especially on intel. Zeroing an entire page of stack with memset (which is what was previously happening) is expensive.

[kkysen]

Why do we need so many ptr casts here? If it's to go from a ref to a ptr, can we use ptr::from_{ref,mut}? They're clearer to read, as the manual ptr cast can be doing other things as well, so it's harder to read.

[nnethercote]

AFAICT this is the single biggest potential perf win in rav1d.

[rinon]

In my testing, it looks like the latest version with less unsafe is equivalent to the unsafe pointer writes with the assert bounds check. Commenting that assert out gains a tiny bit on my i7-1260p, but seems equivalent on my Zen 4 7700X.

[kkysen]

I'm not sure what that use std::ptr in ctx.rs is for but the lf_mask.rs changes all LGTM now.