Speed up CI tests and use CPU-only PyTorch in pip CI

[RussTedrake]

Letter font and dimension tests repeatedly ran expensive convex decomposition, while model conversion tested every Menagerie scene on every CI run. Focus these tests on the behavior each needs to exercise and add timing reports to measure the result.

• Font and dimension tests use bounding-box collision geometry. Dimension cases now assert actual mesh height and depth, and use separate output directories. Dedicated CoACD coverage remains for I, A, C, and D; the collision-mode test has an explicit depth and correct subtest labels.
• Menagerie CI converts Panda and ANYmal B in temporary directories, checks exported mesh/material/texture references, then loads and finalizes the converted models with Drake. These cover STL meshes, includes/defaults, and file-backed textures/materials. The shared package cache is not modified.
• TEST_ALL_MENAGERIE=1 retains the full conversion sweep for Menagerie revision updates. Its smoke-test contract remains conversion success; stronger asset and parser checks apply to the two representative scenes. Some other upstream OBJ files already contain dangling material references.
• All three CI test jobs print the 30 slowest pytest phases and upload JUnit reports. Tests remain serial. Dependency installation and model prefetch are unchanged.

Drake's Menagerie test covers raw MJCF parsing (including expected STL failures), not this repository's conversion pipeline. Local synthetic converter tests remain in place.

Validation and measurements

• Local default suite: 13 tests passed in 11.73s; representative Menagerie conversion/loading 5.19s, font test 0.01s, dimension test 0.03s.
• The final strengthened asset checks also passed in the opt-in full sweep: 13 tests and 48 subtests passed in 64.24s.
• Black, isort, and whitespace checks passed.
• Final CI passed all jobs. Before/after timings compare against the preceding instrumentation-only revision:

Runner	Letter tests before → after	Menagerie before → after	Total job before → after
macOS	55.7s → 5.8s	56.9s → 7.7s	9m42s → 7m48s
Ubuntu Poetry	41.9s → 4.2s	80.6s → 7.3s	11m23s → 9m16s

The targeted tests saved about 99s on macOS and 111s on Ubuntu. Other setup/notebook timings still vary across runners. The unchanged pip test job passed in 9m04s, and lint passed in 1m09s. JUnit reports from both Poetry runners were downloaded to verify these measurements.

Download caches were also evaluated during this PR and removed. Linux restore + installation was 13–25s slower; macOS showed no repeatable gain. The unused experimental caches were deleted. The final change adds no caching machinery.

CI-only CPU wheels

The Linux pip job now selects torch/torchvision from PyTorch's CPU wheel index and retains those exact builds using a temporary constraints file during the published manipulation[all] install. Import checks verify both CPU build tags and the absence of a CUDA build. This does not modify pyproject.toml, poetry.lock, or published dependency metadata; downstream GPU installations remain supported. Poetry CI is unchanged.

CPU-wheel CI passed all jobs. The pip job installed torch 2.14.0+cpu and torchvision 0.29.0+cpu, retained those builds through the manipulation install, and passed all 72 tests with no skips.

Pip CI	Before	CPU wheels
Dependency installation	3m46s	1m55s
Tests	3m47s	3m39s
Total job	9m04s	7m05s

The unchanged Ubuntu Poetry and macOS jobs passed in 8m42s and 8m13s respectively. Workflow YAML/shell syntax checks passed locally. No changes were made to the package manifest or lockfile.

Poetry experiment result

A temporary CPU override for Linux Poetry was also evaluated. It preserved locked versions and original metadata, but full setup took 2m08s versus 2m17s in the prior run: only a nine-second difference. Given runner variability, added re-locking machinery, and macOS remaining above eight minutes, the override was removed. That experiment was superseded after successful setup. The final tree is byte-for-byte identical to the previously passing CPU-pip-only revision 210d981.

---

This change is 

Ubuntu prerequisite cleanup

Ubuntu Poetry CI now skips system Jupyter packages (INSTALL_JUPYTER=0) because its Python environment supplies the notebook dependencies. Local prerequisite installation and pip CI keep their existing default; the pip notebook harness relies on system nbformat/nbconvert. Both Ubuntu jobs use the distribution Git package, retain apt lists until dependency setup completes, and avoid a redundant apt update in the prerequisite script. This preserves clean-container installation coverage while reducing duplicate setup work.

Validation: shell syntax and commit checks passed; final CI is green on all jobs. Ubuntu Poetry finished in 8m00s versus 9m22s on the preceding baseline. Dependency setup fell from 2m51s to 2m06s (the first attempt also measured 2m05s), prefetch from 51s to 29s, and tests from 5m04s to 4m59s. Faster downloads also contributed, so the full 82-second improvement is not attributed solely to the code change. Final pip CI: 6m55s; macOS: 7m40s.

[RussTedrake]

@RussTedrake reviewed 5 files and all commit messages, and made 1 comment.
Reviewable status: 0 of 5 files reviewed, all discussions resolved.