* Add Python 3.12 asyncio compatibility updates

- replace legacy asyncio task scheduling with create_task() in hot paths
- switch loop access to get_running_loop() and explicit event loop setup for Python 3.12+
- move AuthTable auth caches from class state to instance state to avoid cross-connection leakage
- update package metadata to declare Python 3.12+ support

* Add gitignore for Python cache files

- ignore __pycache__/ directories generated by local Python runs

* Add local proxy benchmark script

- compare direct, HTTP, and SOCKS relay throughput on a local echo server
- report request rate, bandwidth, average latency, and p95 latency
- allow running from the repository root without installing the package

* Optimize HTTP and WebSocket header parsing

- share HTTP/WS request-head parsing to reduce repeated decode(), split(), and temporary header dict creation
- reuse filtered header bytes when rebuilding HTTP request lines in accept and channel paths
- benchmark basis: compared current proto.py against commit 53a2919 with tests/benchmark_http_headers.py, 5 runs each, 20000 iterations, 20 extra headers
- measured improvement: http_accept 17.58us -> 16.84us (~4.2% faster), http_channel 17.63us -> 16.85us (~4.4% faster), ws_accept 16.82us -> 15.68us (~6.8% faster)

* Reduce drain frequency in relay loops

- batch relay writes up to 256 KiB before awaiting drain() in generic and HTTP relay paths
- keep forced drain() for response paths that must flush immediately
- benchmark basis: compared current proto.py against commit f888a39 with tests/benchmark_proxy.py, 5 runs each, 400 requests, concurrency 20, payload 16384 bytes, warmup 3
- measured improvement: HTTP relay 9666 req/s -> 9818 req/s (~1.6% faster), SOCKS5 relay 8788 req/s -> 8907 req/s (~1.3% faster), with lower average latency on both paths

* Optimize UDP relay bookkeeping

- replace O(n) udpmap eviction scans with OrderedDict-based LRU tracking in ProxyDirect
- reuse udpmap lookups and centralize touch/discard helpers for UDP connection state updates
- preserve existing round-robin scheduling after rejecting the attempted rr cursor change
- benchmark basis: synthetic UDP bookkeeping microbenchmark comparing previous min(update) eviction against the new LRU path across sizes 30, 100, 300, and 1000
- measured improvement: about +475% at size 30, +480% at size 100, +451% at size 300, and +724% at size 1000

* Reduce stat updater allocation overhead

- replace nested lambda-based stat updaters with ConnectionStat and StatUpdater helper objects
- keep the existing modstat call shape while avoiding repeated closure creation on connection setup
- benchmark basis: synthetic modstat microbenchmark comparing the previous nested-lambda updater against the new object-based updater
- measured improvement: 2189206 ops/s -> 3896353 ops/s (~78.0% faster)

* Group relay tasks with asyncio.TaskGroup

- add a relay_with_taskgroup helper and await paired inbound/outbound relay tasks inside stream_handler
- keep the existing echo and empty fast paths unchanged while grouping the main bidirectional proxy relay lifecycle
- benchmark basis: compared current server.py against commit 67523d9 with tests/benchmark_proxy.py, 5 runs each, 400 requests, concurrency 20, payload 8192 bytes, warmup 3
- validation result: no large regression observed; direct throughput changed from 23336 req/s to 22811 req/s (~-2.3%), HTTP from 9953 req/s to 10436 req/s (~+4.9%), and SOCKS5 from 10022 req/s to 9654 req/s (~-3.7%), with smoke benchmark and compileall both passing

* Reduce byte-processing overhead in protocol hot paths

- decode HTTP admin header blocks once and reuse both the header dict and rendered line text
- switch HTTP request-line detection in http_channel() to the bytes-based matcher to avoid per-chunk decode on the fast path
- avoid extra bytes reallocation in SOCKS domain parsing by splitting the length byte from the host payload
- centralize WebSocket frame masking with xor_mask_bytes() to simplify the byte-processing path used for masking and unmasking
- validation: python3 -m compileall pproxy and python3 tests/benchmark_proxy.py --requests 50 --concurrency 5 --payload-size 4096 --warmup 1 both passed

* Clean up asyncio task imports and Transparent interface

- replace module-level create_task aliases with from asyncio import create_task in proto.py and server.py
- add a default Transparent.query_remote() implementation so static analysis can see the required interface