Improve performance without breaking API

[alanmcgovern]

The original PR was pushed a bit too late and
wasn't reviewed before 1.0 was released. As such,
the changes can't be included now without breaking
API.

Redo some of the changes to gain most of the benefits
without the API breakage.

1. Nothing in the library uses the public mutable
   state defined in QoiCodec.These properties/fields
   have been replaced with immutable/readonly properties
   instead.

2. QoiEncoder has been optimised a bit to leverage
   the performance gains from BinaryPrimitives and Span.
   Also, a Span based overload was added so that images
   could be encoded into a reusable buffer.

3. QoiDecoder has a new overload which receives a
   Span, allowing images to be decoded from anything
   which can create a Span, and also allows for
   multiple images to be decoded from the same underlying
   array.

[alanmcgovern]

@NUlliiON I closed the other PRs and submitted this instead.

Some of the changes from the older PRs are still needed, but at least this one is API compatible with the current 1.0 release.

[alanmcgovern]

Optimised:

Method	Mean	Error	StdDev	Gen 0	Gen 1	Gen 2	Allocated
QoiEncoding	5.273 ms	0.0221 ms	0.0185 ms	164.0625	164.0625	164.0625	2 MB

Method	Mean	Error	StdDev	Gen 0	Gen 1	Gen 2	Allocated
QoiDecoding	3.064 ms	0.0091 ms	0.0085 ms	496.0938	496.0938	496.0938	2 MB

Original:

Method	Mean	Error	StdDev	Gen 0	Gen 1	Gen 2	Allocated
QoiEncoding	5.524 ms	0.0252 ms	0.0224 ms	171.8750	171.8750	171.8750	2 MB

Method	Mean	Error	StdDev	Gen 0	Gen 1	Gen 2	Allocated
QoiDecoding	3.978 ms	0.0794 ms	0.1326 ms	496.0938	496.0938	496.0938	2 MB

NOTE: There are further improvements possible but I want to get a few PRs landed before I start adding new API. Specifically, i want the fix-tests branch landed so I don't have to keep modifying benchmarks to compare.

[Terria-K]

This optimization does have a weird bugs. Comparing these two images, the optimized code has a two extra black pixels while the original code doesn't have one.

Looking at the North West tile textures or the left ledge textures, you can see those two black pixels.

(Original)

(Optimized)

[bwiklund]

This optimization does have a weird bugs. Comparing these two images, the optimized code has a two extra black pixels while the original code doesn't have one.

Looking at the North West tile textures or the left ledge textures, you can see those two black pixels.

I don't see any difference, even putting these on top of each other with a "difference" blend in aseprite. (other than the time and moved crate) are you sure these are the right images?

[Terria-K]

This optimization does have a weird bugs. Comparing these two images, the optimized code has a two extra black pixels while the original code doesn't have one.

Looking at the North West tile textures or the left ledge textures, you can see those two black pixels.

I don't see any difference, even putting these on top of each other with a "difference" blend in aseprite. (other than the time and moved crate) are you sure these are the right images?

if you look closely at the corner, you can see these two black pixels in the optimized version, which shouldn't be appearing.

[bwiklund]

ah, that's probably from the changes in the way the lookup table is initialized. seems like an easy fix.

in any case i don't think the author of this repo is maintaining it anymore, someone (@alanmcgovern ?) should fork it. we'd swap from this repo for the loading time boost for sure.

[bwiklund]

also @alanmcgovern it'd be good to see a benchmark for a large atlas-sized image (like 2k, 4k, 8k)