ci: cache pip wheels across all test jobs

[TrevorBasinger]

macos-15-intel test job takes 15 minutes. ~6.5 of those are "Install dependencies" — and the dominant chunk inside that is ~3.5 minutes spent compiling dependency-injector from source because PyPI doesn't ship a macosx_*_x86_64 wheel for it at the versions we depend on.

Compare:

Linux (ubuntu-latest): install ~33s
macOS arm64 (macos-latest): install ~42s
macOS x86_64 (macos-15-intel): install ~388s ← 6.5 minutes

The macOS x86_64 lane is just slower at test execution too (older Intel hardware), so the test step alone is ~8 min. Not much we can do about that without trimming the macOS test set. The install side is fixable.

actions/setup-python@v5 has a built-in pip wheel cache that backs onto GitHub-hosted cache storage (per-repo, ~10 GB, keyed automatically by OS + arch + Python version + the contents of cache-dependency-path). Once a wheel is built or downloaded once, subsequent runs reuse it.

Adds cache: pip + cache-dependency-path: pyproject.toml to every actions/setup-python step:

• lint
• typecheck
• test (Python 3.10/3.11/3.12/3.13 matrix)
• test-macos (arm64 + x86_64 matrix)
• test-tracer-privileged

Each (OS, arch, Python version) combination gets its own cache entry — no collisions across the matrix. The key invalidates on pyproject.toml changes, so a dep version bump correctly forces a fresh resolve.

Expected impact on cached runs:

• macos-15-intel install: 388s → ~30-60s
• macos-15-intel total: 15m → ~10-11m
• Linux install: 33s → ~5-10s
• Linux total: 4-5m → ~3-4m

First run after this PR merges (or after any pyproject.toml change) still pays the full install cost.