Simplify the Spiral-generated convolutional decoder

[argilo]

Here I've continued cleaning up the convolutional decoder.

Since #755 is not merged yet, I've included it here, and will rebase the PR once it's merged.

In two additional commits I've made the following changes:

1. Remove loop unrolling. Each of the SIMD protokernels had an unrolled loop that processed two bits of input in a row, as well as special processing for the leftover bit, if any. Since the processing step for each bit is large, I don't see any point in unrolling the loop, and removing the unrolling does not appear to have any impact on performance.
2. Remove temporary variables. Spiral generates temporary variables for every single quantity and memory address, which makes the code difficult to understand. I removed these, which again appears to have no impact on performance but makes the code much more readable.

With these changes, the tests still pass. I also verified (on x86 and ARM) that various vector lengths still work:

for v in {0..1024}; do apps/volk_profile -n -R k7_r2 -v $v -i 10 | grep fail; done

[argilo]

Rebased.

[jdemel]

@argilo Thanks! I really appreciate your work here. Before I merge this, how are the test run? We had a lot of issues with this kernel in past. I'd like to make sure, we don't screw this kernel up again.

[argilo]

CI should do a decent job of testing the kernel now, since it verifies that all the protokernels produce identical bit streams when given noise as input. One thing that CI doesn't yet do is test with different vector lengths, which I did like so:

for v in {0..1024}; do apps/volk_profile -n -R k7_r2 -v $v -i 10 | grep fail; done

[argilo]

I ran that test on x86-64, ARM-64, and ARM-32.

[jdemel]

Great. Merging.