fix(python-uv): let UV use its own persistent cache

[bnusunny]

Issue #, if available: #903

Description of changes

UvRunner._ensure_cache_dir() pointed UV's --cache-dir at a uv-cache directory inside the build's scratch directory. Scratch comes from osutils.mkdir_temp() in SAM CLI's ApplicationBuilder._build_function() and is deleted when the build finishes, so UV started from a cold cache on every function of every build. Within one sam build each function re-downloaded shared dependencies (build time scaled linearly with function count), and the next build started cold again. The explicit flag also overrode any UV_CACHE_DIR the user had set, so there was no workaround.

This removes _ensure_cache_dir(). With no --cache-dir, UV uses its own default cache, which persists across functions and builds — matching python_pip, which never sets PIP_CACHE_DIR and inherits pip's user-level cache. A caller-supplied config.cache_dir is still forwarded, and UV_CACHE_DIR is honoured again.

scratch_dir is dropped from install_requirements(), where it had no remaining use. It stays on the manifest handlers: _build_from_lock_file still writes exported requirements there, and _build_from_requirements keeps it to match the signature shared by handlers dispatched from build_dependencies().

Description of how you validated changes

Existing suites: 831 unit, 13 functional, 20 integration — all pass. ruff and black clean.

New unit tests assert no --cache-dir is passed and no cache directory is created by default, that an explicit cache_dir is still honoured, and that a build never derives cache_dir from scratch. I verified these fail against the old implementation on the assertion (not merely on the signature change) by re-introducing the old cache-defaulting logic under the new signature.

Measured with boto3 + pydantic + requests, medians of 5 iterations, all paths on one filesystem, UV 0.11.28:

	before	after (warm)
requirements.txt	2622 ms	245 ms
pyproject.toml	2963 ms	481 ms
4 functions	10436 ms	3236 ms
8 functions	20658 ms	4051 ms

Cold first build is unchanged (2664 ms vs 2622 ms, within noise), so nothing gets slower.

Two risks I checked before proposing a shared cache:

• Concurrency (sam build --parallel): six concurrent installs against one cold cache completed with no errors and identical output trees. UV locks its cache.
• Hardlinks: files installed via --target report nlink=1, so UV copies rather than hardlinks for --target installs and a persistent cache can't be corrupted by later mutation of build artifacts.

Checklist

• Review the generative AI contribution guidelines

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.

[vicheey]

Approved!
One small comment, not blocking: consider documenting the full cache precedence chain (config.cache_dir > UV_CACHE_DIR > UV default) in the DESIGN.md caching section.