More changes:

• option to set a file/folder to load on startup. In the case of a file, the file is simply loaded and starts playback during the nintendo logo. For a folder, the entire folder (non-recursive) is loaded.
• ovl playlist now focuses on the currently playing song.
• ovl browser now auto focuses to the first element, rather than being out of focus until a button is pressed.
• optimised the ovl dir scanning by reading batches of 64 entries at once.
• optimised ovl loadAll sort function by passing the strings by ref, rather than value.
• updated the workflow so that it always builds on both push and pull_request, however it only pushes a new build on push.

• heap was increased to 300k (+100k from last few commits) due to dr_mp3 running out of memory loading certain files. overall we are still using 8k less than the latest release.
• a file limit has been set to the ovl browser in order to prevent against crashes if the user tries to open a very large folder due to running out of memory. i've set the limit to 2k entries, which should be more than enough. i reasoned that every 3 entries is max 3k so 2k entries would be max 682k, which gives us plenty of headroom.
• like above, ive set a limit to the loadAll function in ovl so limit itself to 512 entries, as this is the max amount of songs that can be added to a playlist in sys-tune.

• fixes crash when seeking a wav file due to the source->seek() function being called on the main thread which only has 4k stack, increased the stack to 16k.
• added buffered io. this works like fread where it keeps an internal copy of the last read data, and always reads data in chunks of its internal buffer size unless the data passed into read is bigger than the internal buffer, in that case it's read in place.
• due to the above change, the mp3 chunk size has been reduced to its minimum of 16k.
• due to the above change, the heap has been reduced further because we no longer alloc that extra 48k.
• reduced the latency between songs by making the first sample buffer very small, so it finished reading / decoding a lot faster.
• the decode function now only resamples the data if the src freq != dst freq. sdl does handle this already, however it still has to make a copy of the data, and the decode loop is potentially ran multiple times. so, in the case that resampling isn't needed, decode is called once.
• re-shuffled the shuffle playlist when shuffle mode is enabled. this is similar to how it works on yt music.

I'd be more comfortable if the individual patches only addressed individual issues.
To track the heap usage, we could expose mallinfo in an IPC or shim new/delete.
Maybe we even want to use area-allocators for everything we control and let dr_libs use the rest?