Fix: Add batch upload for prime evals

[d42me]

Note

Introduces adaptive, concurrent batching for pushing evaluation samples in both sync and async clients.

• New push_samples implementations split samples into size-aware batches via _build_batches and upload concurrently (ThreadPoolExecutor for sync, semaphore-limited coroutines for async)
• Configurable limits: max_payload_bytes (default 512KB), max_workers/max_concurrent with validation
• Returns {"samples_pushed": <count>} and aggregates per-batch errors into a single EvalsAPIError
• Helper methods _upload_batch (sync) and _build_batches added; existing create/finalize/list/get/update logic unchanged aside from minor formatting

^{Written by Cursor Bugbot for commit b67f8cc. This will update automatically on new commits. Configure here.}

[cursor]

Concurrent thread access to non-thread-safe httpx client

Medium Severity

The sync push_samples method uses ThreadPoolExecutor with max_workers=4 by default to upload batches concurrently. Each worker thread calls self.client.request() on the shared client instance. The prime_evals.core.APIClient provided by this package uses httpx.Client, which is documented as not thread-safe. Users following the README's documented pattern of using APIClient with EvalsClient would encounter potential race conditions in the HTTP connection pool, which could cause intermittent failures.