perf: share few-shot preamble across prompts via PromptParts

[dan504512]

Fixes #446

Description

Share few-shot preamble across prompts via a new PromptParts dataclass,
reducing batch prompt memory from O(N × preamble_size) to O(1 × preamble_size + N × small_parts).

• Add PromptParts(prefix, examples, suffix) frozen dataclass in prompting.py with __str__() for backward-compatible string conversion
• QAPromptGenerator caches formatted examples in __post_init__ and exposes render_parts() returning PromptParts; render() reimplemented via str(render_parts(...))
• PromptBuilder.build_prompt() and ContextAwarePromptBuilder.build_prompt() return PromptParts instead of str
• _build_request() in gemini_batch.py emits 3 text parts in contents[0].parts when given PromptParts with non-empty examples, keeping the shared reference intact
• Gemini single-prompt, OpenAI, and Ollama providers convert PromptParts to str at entry point (negligible memory impact since they process one at a time)
• GCSBatchCache._compute_hash() resolves non-primitive values (e.g. PromptParts) to str transiently at hash time, producing hashes identical to the old string-based format. Only one temporary concatenated string exists at a time during sequential hash computation, so peak memory is unaffected. Cache compatibility is preserved — no cache misses on upgrade.

Expected impact at batch_length=10000 with ~300 KB examples:
10,000 × 640 KB = 6.4 GB → 1 × 640 KB + 10,000 × ~1 KB = ~10 MB (640× reduction)

Memory benchmarks

Tested with 1000 prompts, ~154 KB examples per prompt:

Version	Batch prompts	key_data_list	Cache compat
main (str only)	150.2 MB	0.4 MB	baseline
This PR	0.2 MB	0.4 MB	hashes match ✓

How Has This Been Tested?

• All existing tests updated and passing (433+ pass across py3.10/3.11/3.12)
• New test_build_prompt_shares_examples_reference and test_context_aware_shares_examples_reference verify the memory-sharing invariant via assertIs (all prompts from the same generator share a single examples string object)
• Verified str(PromptParts) produces byte-identical output to old render() for all 4 cases (examples/no-examples × context/no-context)
• Verified _compute_hash produces identical hashes for PromptParts vs plain str prompts, confirming cache key stability
• All 3 import-linter contracts kept
• Format check passes (pyink + isort)

Checklist:

• I have read and acknowledged Google's Open Source Code of conduct.
• I have read the Contributing page, and I either signed the Google Individual CLA or am covered by my company's Corporate CLA.
• I have discussed my proposed solution with code owners in the linked issue(s) and we have agreed upon the general approach.
• I have made any needed documentation changes, or noted in the linked issue(s) that documentation elsewhere needs updating.
• I have added tests, or I have ensured existing tests cover the changes.
• I have followed Google's Python Style Guide and ran pylint over the affected code.

[github-actions]

⚠️ Branch Update Required

Your branch is 1 commits behind main. Please update your branch to ensure CI checks run with the latest code:

git fetch origin main
git merge origin/main
git push

Note: Enable "Allow edits by maintainers" to allow automatic updates.

[github-actions]

⚠️ Branch Update Required

Your branch is 6 commits behind main. Please update your branch to ensure CI checks run with the latest code:

git fetch origin main
git merge origin/main
git push

Note: Enable "Allow edits by maintainers" to allow automatic updates.