SourceT's readFile uses an innefficient chunk size

4096 is not the most optimal chunk size, and so this function is slower at reading files than it could be.

In a quick benchmark, using the example from https://docs.servant.dev/en/stable/cookbook/basic-streaming/Streaming.html, and a 1 gb "README.md" file, curl to /dev/null takes 1.84 seconds.

I was able to optimize that to 0.68 seconds. My optimisation happened to read it into a lazy ByteString and then stream out the toChunks, which is ugly and I don't recommend. (I was actually just investigating how to stream a lazy ByteString through servant when I noticed this performance problem.)

Probably the right fix is to change 4096 to some more optimal chunk size. Data.ByteString.Lazy.Internal.defaultChunkSize is a good one.

Locate SourceT's readFile implementation and inspect how it chooses the 4096-byte chunk size. Compare the current behavior with Data.ByteString.Lazy.Internal.defaultChunkSize using the linked streaming example or an equivalent benchmark. Done means the read path uses an evidence-backed chunk size without changing streaming behavior.

Written by the indexing model from the issue text.