feat: expose async stream cancel handle

[MukundaKatta]

Summary

• return a cancelable async iterator handle from chat.stream() while preserving existing async for usage
• expose cancel() and aclose() so callers can stop the underlying gRPC stream early
• document the early-stop pattern and add a regression test for transport cancellation

Closes #142

Tests

• uv run pytest tests/aio/chat_test.py -q
• uv run pytest tests/aio/chat_test.py::test_stream_cancel_cancels_underlying_rpc tests/aio/chat_test.py::test_streaming tests/aio/chat_test.py::test_stream_creates_span_with_correct_attributes -q
• uv run ruff check src/xai_sdk/aio/chat.py tests/aio/chat_test.py
• uv run pyright src/xai_sdk/aio/chat.py tests/aio/chat_test.py
• uv run ruff format --check src/xai_sdk/aio/chat.py tests/aio/chat_test.py
• git diff --check

[kalyanamdewri]

Same-seed comparison of #143 and #148

I ran fixture seed 142 against main 4358bc235e8641ba5f0cb54599675d098385d4bf, #143 9b55aa6b9b8fa24756a6cacddf91dcbe4bfe5523, and #148 7ed05b5f9e0247af49f1aa24abfd77ec30b4b597.

Ref	Cancel outcome	Fragments delivered	Final gRPC status	Live tasks/channels after 500 ms
main	aclose RuntimeError: asynchronous generator is already running; blocked read PENDING	3	NOT_TERMINAL	1/1
#143	aclose returned; blocked read CancelledError	3	CANCELLED	0/0
#148	aclose returned; blocked read CancelledError	3	CANCELLED	0/0

Under this fixture, #143 and #148 preserved the measured invariant that cancellation converges to a clean terminal state without dangling tracked replay tasks or calls. All 3/3 independent repetitions per ref produced the same normalized result within that ref.

Harness, trace, and environment lock (harness commit 9655539bce02fd5aaa1ad41bf0e50cecf22dd6fa): public evidence gist (report sha256:8f3f2a21eef1bc2e733fadf59dd0c703c5589a9822bfcb192e76deb3e199912d; trace sha256:7284d81230894c360c0b5b4252e3b6145b93736225c4fe3bf8a311a818383bf5). Untested: the live xAI service, real gRPC/TCP channel ownership, SDK retries or retry backoff, and batch streaming. #143 is currently conflicting with main, so this is exact-head behavioral evidence, not integration evidence.