Rather than applying the HF scale to the IRs necessitating them to be truncated
along with increasing the IR size, it can be applied to the input signal for
the same results. Consequently, the IR size can be notably shortened while
avoiding the extra truncation. In its place, the delayed reversed all-pass
technique can still be used on the input for maintaining phase when applying
the bandsplit/hfscalar filter to the input signal.
This update removes the 16/24-bit sample type enum, now always being 24-bit
(other than a very small size saving, there's no practical benefit to storing
16-bit samples). This also reverses the field storage, so no on-load fixup is
needed, and stores the IR delays with 2 bits of sub-sample precision, allowing
for slightly better timing (after resampling, blending, etc).
The SIMD mixers may use a few more elements beyond the irSize, so not limiting
the blend to that size can slightly improve the quality for resampled HRIRs.
Though fine in theory, an issue arises with extra phase interference since the
frequency phases aren't aligned for each response. It would be better to do the
blending before minimum phase reconstruction, where it can blend just the
frequency magnitudes, essentially allowing makemhr to increase the resolution
of the dataset.
Since the delays get bilinearly filtered before use, maintaining a bit of extra
precision can help accuracy even if it ultimately gets rounded to the nearest
integer before use. This should help with resampled HRTFs, which need to scale
the delays that are already rounded, and can also be leveraged by an updated
MHR format.
