feat(gcp-sphinx-docfx-yaml): optimize performance

[daniel-sanche]

Previously, docfx would take > 3 hours on large PR. The compute package alone would take > 70 minutes, so this wasn't a problem that could be solved with sharding

This PR attempts to speed up the task by adding optimizations to the gcp-sphinx-docfx-yaml package:

• 2bb74d8: sort known ids once, instead of inside a loop
• 7a1c741: immediately skip over words without "." characters, since they can't be uuids
• c54391e: use custom no-op StandaloneHTMLBuilder class, to avoid doing rendering work for unused html files
• 7b9ab23: use in-process markdown builder, instead of spanning new process and starting from scratch
• 8dedfb7: cache default settings, instead of re-constructing on each node visit
• 7b91c2d: cache class line numbers, instead of re-building for each one
• 9effc20: disable unused sphinx hooks
• 28478f8: prefer native yaml parser when available. Fall back to pure-python when needed
• 6d9bb07: added fast-path to skip non-google symbols

I had gemini run a comparison between these outputs, and outputs from the previous version, and it saw identical outputs

A follow-up PR re-adds docfx as a sharded pre-submit, with all jobs finishing < 15 mins

[gemini-code-assist]

Code Review

This pull request optimizes cross-reference conversion by skipping plain words that do not contain dots and hoisting the sorting of known_uids out of a loop. However, the changes introduce a critical NameError because hard_coded_references is referenced in find_uid_to_convert where it is not in scope. The review feedback suggests reverting this reference and instead combining the collections as immutable tuples within convert_cross_references to safely avoid shared state mutation.