Well, it's easier to read in a hex editor and seemed like the right thing to do for a "wire format".
The only places where it matters is the header's width/height and order of RGB(A) data. Storing RGBA as ABGR is not a great choice, imho. If you want to be endian independent, you have to read one byte at a time anyway (or swap).
I don't believe it would imply that, because those are independent elements. RGBA would still be stored as R, G, B, A unless you're considering each pixel as a single 32-bit integer. A C structure for the colors would even live in memory in the order you expect. Just like a string is still stored byte-by-byte even in little-endian contexts.
LE vs BE would only affect multi-byte integers and floats, not arrays or structures of single-byte elements.
Because almost everywhere you would use data will expect it as interleaved. Separating the channels would require a second pass to re-interleave the data.
Support for 16/32b and fp-formats would've been nice -- the compressor is actually orthogonal to the bit width of the channels. Custom fp-compression usually just unvolves moving the sign bit into the LSB, and delta encoding the mantissa separately from the exponent -- although the latter isn't as important.
If I had to nitpick something, I'd say choosing all-big-endian is kinda odd in 2021 but it's probably not a biggy even for something like this.