In some cases we can't quantize the floating point data because the
range of the data is unknown. While it's possible to use
meshopt_quantizeFloat to reduce the precision and gain some compression
back, this is often insufficient and suboptimal.

For inputs that represent a vector in 3D space, such as a position or
scale, a good alternative is to use a shared-exponent encoding - it's a
reasonable assumption that we are content with the same (absolute)
precision in all three components.

To be able to encode in shared exp, we use a modified floating point
like format, where we store a 24-bit signed integer mantissa (without
implicit 1) and a 8-bit exponent. This is less precise than a floating
point number - we lose 1 bit - but we gain an ability to individually
select the exponent and mantissa at any level of desired mantissa
precision. Additionally this moves exponent into a single byte, and
stores the mantissa as a two-complement integer - both of these are much
friendlier for vertex codec than a basic float encoding.

While ideally the shared exponent would be stored just once, this
complicates the SIMD decoding and is actually redundant if the output of
the filter is compressed with vertex encoder *and* a general purpose LZ,
because the stream of exponent bytes will be exactly the same between
all three components.

The resulting decoder runs at ~13 GB/s using WASM SIMD and ~2.5 GB/s
using scalar WASM.