[Tracking] Benchmark whether intermediate-aux-file caching is worth implementing

[nicklambourne]

Background

Tectonic, like classical LaTeX, runs multiple passes over a document
when it has cross-references, bibliography, or tables of contents.
Each pass reads a .aux file written by the previous pass.

Bazel currently caches the action output (the PDF + synctex). It
does not separately cache the intermediate .aux / .toc / .bbl
files between passes within a single action — Tectonic does that
internally in its temp dir.

The hypothesis is: for multi-pass documents with stable
bibliographies, we might be able to skip a pass by caching the
.aux file as a Bazel action output of an earlier action, and
feeding it back as an input to a later one.

Why this is probably never worth doing

• Tectonic is fast. A full from-scratch compile of a thesis-sized
  document with biblatex + biber is ~3-5 seconds.
• Bazel's action-output caching already kicks in for the
  no-source-change case: rebuild is ~0.05s, dominated by Bazel
  overhead.
• The savings would only materialise when a user edits a
  non-bibliography source file in a way that changes the body but
  not the citations — a real but narrow case.

What "done" looks like

Either:

1. A benchmark that demonstrates an actually-painful slowdown on a
   real document, with a sketch of a fix.
2. A documented benchmark showing the savings are negligible, with
   this issue closed.

Source: DESIGN.md §5.5.