Resample

get rid of stbir__inline for some functions that don't need it

replace int(floor()) with int(), since that produces same results for non-negative values and is much faster on 32-bit x86

because of we divide by /255 not /256, we lose numerical precision, so can't guarantee that box filter results round to same results as naive

cache filter-width derived computations that use floor/ceil because those were showing up like crazy on VC6 profile despite only being per-scanline-ish?!?
unroll inner loop in horizontal_resample

tweak filter type documentation

Optimizations: Skip zero-coefficient contributions. Reduce the size of the coefficients array when downsampling.

rename STBIR_FLAG_PREMULTIPLED_ALPHA to STBIR_FLAG_ALPHA_PREMULTIPLIED so that both flags have "ALPHA" first

    1. table stored threshhold of transition from i to i+1, but wants to be i-1 to i
    2. table was computed by dividing uchar by 256.0 instead of 255.0, causing it to be 100% wrong

fix some unsigned/signed comparisons;
avoid round() since it's not in pre-C99 C;
remove MAX_CHANNELS since I never ended up needing it;
rename STBIR_EPSILON to STBIR_ALPHA_EPSILON;
don't use STBIR_ALPHA_EPSILON on float input (can't remove it properly due to numeric precision, and they can do it themselves);
optimize subtraction of STBIR_ALPHA_EPSILON;
sorry i forgot to commit these separately!;

Use an even smaller epsilon. Allow the user to avoid the epsilon business if they want to preserve their alpha values.

Add the epsilon value into the alpha instead of using an if statement. It's a tiny bit faster and it can be removed afterwards.

Use the info structure to keep track of our memory block sizes to reduce duplicated code and cut down on errors.