feat: add asyncio support with HTTPX

[Joseph-Mutua]

Summary

• replace the Requests transport with persistent httpx.Client / httpx.AsyncClient transports while preserving the existing synchronous serpapi.Client and module-level helpers
• add serpapi.AsyncClient support for search, archive, account, locations, image upload, concurrent calls, and async pagination
• add explicit close() / aclose() and sync/async context managers, shared parameter behavior, redirects, timeout and error translation, and compatibility handling for the existing per-call proxy/TLS options
• document pip3 and uv installation, async lifecycle and concurrency, and include a deterministic local transport benchmark
• align package metadata with HTTPX and the existing CI matrix by requiring Python 3.8+

Fixes #27

Compatibility notes

The synchronous call pattern remains unchanged. timeout=None keeps the SDK's historical no-timeout behavior, redirects remain enabled, pagination URLs retain existing query parameters, and SerpApi exceptions retain status_code, error, request, and response details.

HTTPX configures proxy and TLS settings at the client level. Constructor-level proxy, verify, and cert settings therefore reuse the connection pool; legacy per-call proxies, verify, cert, and stream arguments remain accepted through a scoped compatibility path.

HTTPX 0.28 requires Python 3.8+, which matches this repository's current Python 3.8-3.14 CI matrix.

Verification

• 119 passed, 42 deselected in the deterministic offline suite
• 34 passed for the transport/client suite on Python 3.8.20 with HTTPX 0.28.1
• Ruff format and lint checks pass with a Python 3.8 target
• Sphinx HTML build passes with warnings treated as errors
• wheel and sdist build successfully and pass twine check
• clean Python 3.8 wheel install exposes Client and AsyncClient and depends on HTTPX, not Requests

Live API tests were not run locally to avoid consuming account searches; the existing CI workflow can exercise those with the repository secret.

Controlled benchmark

20 requests against a local threaded endpoint with a 50 ms response delay, using the median of three runs and excluding client setup/shutdown:

Client	Median	Observed range
requests.Session sequential	1.295 s	1.091-1.338 s
serpapi.Client sequential HTTPX	1.105 s	1.076-1.121 s
serpapi.AsyncClient concurrent	0.122 s	0.121-0.160 s

That is a 9.05x throughput improvement for this concurrent I/O workload. It does not imply that a single search becomes faster.

Review context

@adarshdigievo @jvmvik - I followed the separate AsyncClient direction discussed in #27. The two decisions I would especially appreciate confirmation on are raising the declared minimum to Python 3.8 and retaining legacy per-call proxy/TLS options via scoped clients.