BRR decoding improvements and some cleanup

[PhoenixBound]

This PR combines some general refactoring of brr.c with a new feature that fixes the way instruments 17 and 19 sound. It's not really cleaned up yet but I wanted to create a PR before the week ended. Sorry about that.

• The first commit is mostly a reorganization of brr.c, though it adds some additional error checking in some places to prevent undefined behavior. It also changes the calculation a bit to be closer to "bit-perfect"; I figured this would result in cycles being shorter, since there are fewer possible values for a sample to have (LSB is always clear).
• The second commit adds the main feature: at the end of a sample, EBMusEd now checks to see if the sample would sound different when the first block is re-decoded, and if it does, it tries to decode the loop again and again until it finds a cycle. Many looping samples are apparently affected by this, to varying degrees (some samples take around 10 loops before they settle into a cycle), but the noise samples were hit hardest.
• The third commit changes a strategy introduced in the first commit, so there's slightly less confusing pointer arithmetic. It also should match what emulators (and probably real hardware?) do a little better.

Included in the first commit is code to handle failure for malloc, and the second commit includes realloc as well. There's a to-do comment left in the code about what to do if realloc fails. Right now, it just treats it the same way as the case where a cycle couldn't be found: it creates a loop segment with the last decoded loop portion of the sample. I'm interested in different ideas for what to do there, including whether doing nothing (like the rest of EBMusEd's code) is preferred.