[codex] Stream IMAP fetch literals

[mpscholten]

Summary

Fixes the high memory usage reported in #39 when fetching full IMAP message bodies.

This changes the IMAP fetch path so literal response bodies are read as ByteStrings directly from the stream instead of being parsed through the existing packrat/String parser and then packed back into ByteString.

Root Cause

fetch returned a ByteString, but internally it went through fetchByString, causing large IMAP literals to be represented as:

• the full buffered server response
• packrat parser derivations per byte
• a Haskell String
• a re-packed ByteString

That made full-message fetches allocate far more memory than the message size.

Changes

• Added fetchByByteString and fetchByByteStringR.
• Rewired fetch, fetchPeek, fetchHeader, fetchR, and related APIs through the ByteString-native path.
• Kept fetchByString and fetchByStringR as compatibility wrappers.
• Stream FETCH responses directly from BSStream and read {n} literals with bsGet n.
• Added regression coverage for raw bytes and a 1 MB literal fetch.

Validation

• nix shell github:NixOS/nixpkgs/nixos-25.05#ghc github:NixOS/nixpkgs/nixos-25.05#cabal-install -c cabal test imap-parsers
• 21 tests passed on upstream master.

A compiled synthetic benchmark showed much lower peak RSS after streaming:

• 2 MB literal: about 15 MB max RSS
• 10 MB literal: about 23 MB max RSS